Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS20110060738 A1
Publication typeApplication
Application numberUS 12/646,916
Publication date10 Mar 2011
Filing date23 Dec 2009
Priority date8 Sep 2009
Also published asCA2713507A1, EP2306344A1, US8620919, US20120233167, WO2011031661A1
Publication number12646916, 646916, US 2011/0060738 A1, US 2011/060738 A1, US 20110060738 A1, US 20110060738A1, US 2011060738 A1, US 2011060738A1, US-A1-20110060738, US-A1-2011060738, US2011/0060738A1, US2011/060738A1, US20110060738 A1, US20110060738A1, US2011060738 A1, US2011060738A1
InventorsPatrick Gates, Jeremy Werner, Andrew H. Vyrros, John Andrew McCulloh, Richard Frederick Wagner, Erik Danforth Strahm
Original AssigneeApple Inc.
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Media item clustering based on similarity data
US 20110060738 A1
Abstract
Methods and arrangements for facilitating generation of media mixes for a program participant based at least in part on media library inventory information provided by a number of program participants. Those individuals that decide to be program participants are interested in organizing, maintaining and playing their music, based at least in part, on data derived from a population of other participants in the program. A program participant must send, and the system, receive, data representative of that program participant's media inventory. The system or program determines a relative similarity of each item from the collection of program participants as compared to each other item and from the similarity information clusters of similar items are identified. The clusters can be used to identify clusters of similar items in an individual program participant's media library and therefrom mixes of similar media items can be created.
Images(20)
Previous page
Next page
Claims(31)
1. A computer implemented method for clustering similar media items in a program participant's media library comprising:
dividing a collection of media items that each exist in at least one program participant's media library of a population of program participants into canopies comprising respective groups of potentially similar media items;
recursively sub-dividing the media items in each canopy, each recursion producing non-overlapping, increasingly refined, server clusters of media items determined to be similar to other items in a respective server cluster based on a cluster analysis of similarity data derived from the population of program participants;
terminating the recursive sub-dividing responsive to determining that the server clusters produced from the latest recursion each comprise no more than a determined number of media items, resulting in a plurality of lowest-lowest server clusters; and
creating one or more media item clusters for a particular program participant from items present in that program participant's media library, comprising:
determining respective participant media item clusters for media items found in that program participant's media library from their membership in any of the lowest-level server clusters; and
recursively agglomerating participant clusters that share a common parent server cluster, and which have fewer than a participant cluster minimum number of media items, into respective larger participant clusters until each of those larger participant clusters comprises at least the participant cluster minimum number of media items.
2. The method of claim 1, wherein the canopies divide the program participants' collective media library inventories into groupings based on editorial data from which it can be concluded that media items in different groupings can never be considered similar.
3. The method of claim 2, wherein the editorial data comprises genre information.
4. The method of claim 1, wherein the cluster analysis comprises a k-means analysis.
5. The method of claim 1, wherein the media items in each canopy are grouped by artist before the cluster analysis.
6. The method of claim 1, wherein the determined number of media items in a lowest-level server cluster is 1000 or less.
7. The method of claim 1, wherein the determined number of media items in a lowest-level server cluster is 100 or less.
8. The method of claim 1, wherein the determined number of media items in a lowest-level server cluster is 50 or less.
9. The method of claim 1, wherein the participant cluster minimum number of media items comprising is equal to or greater than 20 items.
10. The method of claim 1, wherein the participant cluster minimum number of media items comprising a participant cluster is equal to or greater than 50 items.
11. The method of claim 1, wherein the similarity data derived from the population of program participants comprises a vector space, wherein each vector within the vector space represents a media item and all media items similar to the media item.
12. A method of creating mixes of media items from a program participant's collection of media items comprising:
sending to a server, information describing individual media items of a program participant's collection of media items;
receiving from the server, data describing clusters of media items contained within the program participant's collection of media items, the clusters having been compiled based on an agglomeration of tracks found in hierarchically related clusters of media items in a media item inventory available to the server, the clusters of media items in the inventory having been determined based on a cluster analysis of similarity data derived from a population of program participants; and
determining a mix of media items, the mix comprising media items selected from one of the clusters of media items contained within the program participant's collection of media items received from the server.
13. The method of claim 12, wherein the mix of media items comprises a playlist generated from songs selected from one of the clusters and songs similar to the songs selected from one of the clusters.
14. The method of claim 12, wherein the songs comprising the mix of media items are chosen based on a degree of similarity between each of the songs and the one cluster.
15. The method of claim 12, wherein the similarity data derived from the population of program participants comprises a vector space, wherein each vector within the vector space represents a media item and all media items similar to the media item.
16. A device comprising:
a media storage unit configured to store data describing clusters of media items stored in a program participant's collection of media items, the clusters having been compiled based on an agglomeration of tracks found in hierarchically related clusters of media items available from a server inventory, the clusters of media items in the server inventory having been determined based on a cluster analysis of similarity data derived from a population of program participants;
a tangible computer readable medium storing instructions for configuring a processor to select a mix of media items selected from media items identified in one of the clusters of media items in the program participant's collection of media items; and
a processor to be configured by instructions obtained from the computer readable medium.
17. The device of claim 16 further comprising:
a communications interface configured to send to the server, information describing individual media items contained with a program participant's collection of media items, and further configured to receive, from the server, the data describing the clusters of media items in the program participants collection of media items.
18. The device of claim 17, wherein the cluster data comprises information identifying a plurality of clusters and media items identified within each of the plurality of clusters.
19. The device of claim 17, wherein the mix comprises a list of similar media items generated from the cluster data.
20. A system for clustering similar media items in a program participant's media library comprising:
a server configured to recursively sub-divide a collection of media items formed from the media items that exist in at least one program participant's media library of a population of program participants, each recursive sub-division producing non-overlapping, increasingly refined, server clusters of media items determined to be similar to other items in a respective server cluster based on a cluster analysis of similarity data derived from the media libraries of the population of program participants and to terminate the recursive sub-dividing responsive to a determination that the server clusters produced from the latest recursive sub-division each comprise no more than a determined number of media items;
a server configured to create one or more media item clusters for a particular program participant from items present in that program participant's media library based on the server clusters, by performing a method comprising: determining participant clusters for media items from that program participant's media library that are found in one or more lowest-level server cluster, and recursively agglomerating, into larger participant clusters, participant clusters that share a common parent server cluster until each of those larger participant clusters comprises at least a determined minimum number of media items; and
a media playing device configured to generate a mix of similar media items for playback based on a selection of media items represented in one of the participant clusters.
21. The system of claim 20, wherein the servers make up part of one server array.
22. The system of claim 20, wherein the servers comprise the same server.
23. The system of claim 20, wherein the similarity data derived from the population of program participants comprises a vector space, wherein each vector within the vector space represents occurrences of an individual media item in each program participant's media library.
24. The system of claim 20, wherein the similarity data derived from the population of program participants comprises a similarity matrix representing incidences of co-occurrences of an individual media items in among program participants' media libraries.
25. A machine-readable medium having stored thereon machine-readable instructions for causing a machine to perform a method comprising:
recursively sub-dividing a collection of media items that each exist in at least one program participant's media library of a population of program participants, each recursion producing non-overlapping, increasingly refined, server clusters of media items determined to be similar to other items in a respective server cluster based on a cluster analysis of similarity data derived from the population of program participants;
terminating the recursive sub-dividing responsive to determining that the server clusters produced from the latest recursive sub-division all comprise no more than a determined number of media items; and
creating one or more media item clusters for a particular program participant from items present in that program participant's media library, comprising:
forming respective participant clusters of media items found in that program participant's media library based on membership of those media items in respective lowest-level server clusters; and
recursively agglomerating participant clusters that share a common parent server cluster into larger participant clusters until each of those larger participant clusters comprises at least a determined minimum number of media items.
26. The machine-readable medium of claim 25, wherein the method further comprises
dividing the collection of media items that each exist in at least one program participant's media library of a population of program participants into canopies comprising respective groups of potentially similar media items, wherein items in different canopies cannot be divided into the same server cluster.
27. The machine-readable medium of claim 26, wherein the canopies are determined based on editorial data.
28. The machine-readable medium of claim 27, wherein the editorial data comprises genre information.
29. The machine-readable medium of claim 25, wherein the cluster analysis comprises a k-means analysis.
30. A machine-readable medium having stored thereon machine-readable instructions for causing a machine to perform a method comprising:
sending to a server, information describing individual media items contained with a program participant's collection of media items;
receiving from the server, data defining clusters of the media items contained within the program participant's collection of media items, the clusters having been compiled based on an agglomeration of tracks found in hierarchically related clusters of media items in a server inventory, the clusters of media items in the server's inventory having been determined based on a cluster analysis of similarity data derived from a population of program participants; and
determining a mix of media items, the mix comprising media items represented in one of the clusters of media items contained within the program participant's collection of media items received from the server.
31. The machine-readable medium of claim 30, wherein each of the clusters identifies a determined minimum number of items, and the cluster was formed by agglomerating increasingly dissimilar items until the cluster identified the minimum number of items.
Description
    CROSS REFERENCE TO RELATED APPLICATIONS
  • [0001]
    This application claims the benefit of U.S. Provisional Patent Application No. 61/240,630, entitled “MEDIA ITEM CLUSTERING BASED ON SIMILARITY DATA”, filed on Sep. 8, 2009, and which is hereby expressly incorporated herein by reference in its entirety.
  • BACKGROUND
  • [0002]
    1. Technical Field
  • [0003]
    The present disclosure relates to clusters of media items and more specifically to generating media mixes based on clusters of media items.
  • [0004]
    2. Introduction
  • [0005]
    Before the age of compressed media files such as MPEG, ACC, WMA media was stored on magnetic tapes and later compact discs (CDs). Especially with reference to music media files, consumers typically bought music media as albums comprising several music tracks. Frequently these albums contained tracks which a listener would associate with varying degrees of enjoyment, either because of the suitability of the individual track to the listener's tastes or the listener's present listening mood.
  • [0006]
    This frequently led to the situation where the tracks a listener would like to listen to were scattered over several or even many albums. Motivated by this situation, user's commonly created “mixes” of their favorite audio tracks or “mixes” of audio tracks that were similar and therefore sounded great when played together on the same tape or compact disc.
  • [0007]
    To create these mixes, users had to have great knowledge of all of the tracks in their collection. Further users had to record each track from its original media onto another media. Often user's spent hours listening to tracks to determine which tracks could fit on the limited space of a tape or CD and to determine the order.
  • [0008]
    Today many media consumers buy individual tracks rather than buy complete albums. Also users today often listen to a greater number of tracks in their library since their entire collection can fit on one device instead of many tapes or CDs.
  • [0009]
    Despite this relative convenience of having access to every track in their media library at any time and being able to access each track with a scroll and a click, listener's still create mixes of tracks, often in the form of playlists. However, creating these mixes is still time consuming processes requiring a great deal of user interaction and many user decisions. Accordingly there is a need to eliminate the burden on the listener while still creating high quality mixes.
  • SUMMARY
  • [0010]
    Additional features and advantages of the concepts disclosed herein are set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the described technologies. The features and advantages of the concepts may be realized and obtained by means of the instruments and combinations particularly pointed out in the appended claims. These and other features of the described technologies will become more fully apparent from the following description and appended claims, or may be learned by the practice of the disclosed concepts as set forth herein.
  • [0011]
    The present disclosure describes methods and arrangements for identifying clusters of media items in a program participant's media library for use in creating mixes of media items. To create the highest quality mixes, the clusters are based on the input of an entire population of users.
  • [0012]
    A population of program participants all share information about the contents of the individual media items in their music libraries with a server hosting the program in exchange for receiving clustering data related specifically to their individual music libraries.
  • [0013]
    In some embodiments items are identified as similar based on an analysis of the program participants' libraries. For every media item represented in at least two program participants' media libraries a vector is plotted representing the presence of that media item in the respective program participants' libraries. Each vector, represents a media item and can be compared to every other vector. For any two items, the closer the relationship between any two vectors, the more similar those two items are said to be with respect to each other.
  • [0014]
    Using the similarity data identified above, a clustering analysis can map the individual media items collected from the program participants' libraries into server clusters or groupings of media items in the server database that are all relatively similar to each other. In some embodiments, the clustering analysis can be applied recursively on each server cluster to result in smaller server clusters comprising media items having a stronger similarity relationship to each other than the media items comprising the “parent” cluster. By recursively applying the cluster analysis a tree-like relationship of server clusters can be generated with each “parent” server cluster having two or more “child” server clusters dividing the contents of the “parent” server cluster among the “child” server clusters. This recursive analysis can be applied as often as desired, but in some embodiments the analysis is reapplied until “child” server clusters have 1000 media items or less, or 100 media items or less, or 50 media items or less. Server clusters meeting this criterion will be the lowest-level server clusters in the tree.
  • [0015]
    In some embodiments the clustering analysis is combined with editorial partitions. While it is simplest to perform the editorial partitions before the clustering analysis, the editorial partitions can be performed at any time. In some embodiments canopies of non-overlapping groups of potentially similar media items are created based on genre information before the clustering analysis is performed. For example, in the case of music, music media items that belong to the “rap” genre will never be considered similar to music media items that belong to the “opera” genre. Accordingly, the entire collection of media items can be separated into large canopies of non-overlapping groups such as “rock”, “rap”, “classical”, “opera”, etc. However, just because media items are grouped together into the same canopy, it does not mean that each of those items is similar to each other, but they can be.
  • [0016]
    The server can generate participant clusters that contain only media items present in a program participant's media library by projecting the similarity relationships represented by the server clusters into a program participant's media library. Since the server already knows of the contents of a program participant's media library, the server can determine which lowest-level server clusters contain each item in the program participant's media library and form participant clusters therefrom.
  • [0017]
    If any one of the participant clusters contains too few media items, for example less than 20 or less than 50 media items, the server can recursively agglomerate clusters sharing a common parent server cluster. By repeating the agglomeration action, a program participant's media items that descend from a common parent are accumulated into a larger participant cluster until the participant cluster has reached an acceptable size. In some embodiments the agglomeration action continues until only a determined number of clusters remain.
  • [0018]
    The server can send, and a client device can receive, information about participant clusters created from the media items of a program participant's media library. The client device can receive a listing of all media items contained in any given cluster from the server. As described above, the clusters have been compiled based on an agglomeration of the program participant's tracks found in hierarchically related participant clusters of media items on the server. The clusters of media items have been determined based on a cluster analysis of the media items in the server's inventory based on the relative similarity of each song in the inventory to each other, the relative similarity among songs have been determined by an analysis of the incidences of co-occurrence of media items within different program participants' media libraries.
  • [0019]
    The resulting participant clusters can be used to generate and play a mix of media items. The mix comprises media items represented in one of the clusters of media items contained within the program participant's collection of media items received from the server. In some embodiments, the mix contains a selection of some of the items represented in a cluster and other songs deemed similar to those media items. In other words, the media items of the cluster can act as seed items for the rest of the mix. It will be appreciated that there are numerous ways to select which media items from within the cluster will be used as seed items to create the mix. For example in some embodiments, the media items can be selected based on their relative similarity to the cluster as a whole. In some embodiments, all the items of the cluster can be used in creating the mix.
  • [0020]
    It should be appreciated that unique aspects of this overall program or system take place on servers of the media supplier (online music store), the media playing devices of the participants, and often intermediate facilities, such as the users' personal computer that accommodates communication between the playing device and online store's computer systems.
  • BRIEF DESCRIPTION OF THE DRAWINGS
  • [0021]
    In order to best describe the manner in which the above described embodiments are implemented, as well as define other advantages and features of the disclosure, a more particular description is provided below and is illustrated in the appended drawings. Understanding that these drawings depict only exemplary embodiments of the invention and are not therefore to be considered to be limiting in scope, the examples will be described and explained with additional specificity and detail through the use of the accompanying drawings in which:
  • [0022]
    FIG. 1 illustrates an example system embodiment;
  • [0023]
    FIG. 2 illustrates an example method embodiment for receiving and identifying items in a program participants library and generating similarity data;
  • [0024]
    FIGS. 3A and 3B illustrate a method of determining correlation data;
  • [0025]
    FIG. 4 illustrates a sample purchase matrix;
  • [0026]
    FIG. 5 illustrates a column in a master similarity table;
  • [0027]
    FIG. 6 illustrates an individual similarity table;
  • [0028]
    FIG. 7 illustrates an exemplary method of identifying clusters in a vector space;
  • [0029]
    FIG. 8 illustrates an exemplary method of identifying sub-clusters in a vector space;
  • [0030]
    FIG. 9 illustrates an example server cluster embodiment;
  • [0031]
    FIG. 10 illustrates an exemplary method of generating server clusters;
  • [0032]
    FIG. 11 illustrates an exemplary method of generating server clusters;
  • [0033]
    FIG. 12 illustrates an exemplary method of computing a cluster vector;
  • [0034]
    FIG. 13 illustrates an example server clusters and canopies embodiment;
  • [0035]
    FIG. 14 illustrates an exemplary method of generating participant clusters;
  • [0036]
    FIG. 15 illustrates an exemplary method of generating participant clusters;
  • [0037]
    FIG. 16 illustrates an example system embodiment;
  • [0038]
    FIG. 17 illustrates an exemplary method of receiving similarity data;
  • [0039]
    FIG. 18 illustrate an example playlist method embodiment;
  • [0040]
    FIG. 19 illustrates an example media mix method embodiment;
  • [0041]
    FIG. 20 illustrates an exemplary method embodiment for recommending items for purchase;
  • [0042]
    FIG. 21 illustrates an example system embodiment; and
  • [0043]
    FIG. 22 illustrates an example system for synching media and similarity tables to a digital media player.
  • DETAILED DESCRIPTION
  • [0044]
    Various embodiments of the disclosed methods and arrangements are discussed in detail below. While specific implementations are discussed, it should be understood that this is done for illustration purposes only. A person skilled in the relevant art will recognize that other components, configurations, and steps may be used without parting from the spirit and scope of the disclosure.
  • [0045]
    With reference to FIG. 1, an exemplary system includes a general-purpose computing device 100, including a processing unit (CPU) 120 and a system bus 110 that couples various system components including the system memory such as read only memory (ROM) 140 and random access memory (RAM) 150 to the processing unit 120. Other system memory 130 may be available for use as well. It can be appreciated that the system may operate on a computing device with more than one CPU 120 or on a group or cluster of computing devices networked together to provide greater processing capability. The system bus 110 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. A basic input/output (BIOS) stored in ROM 140 or the like, may provide the basic routine that helps to transfer information between elements within the computing device 100, such as during start-up. The computing device 100 further includes storage devices such as a hard disk drive 160, a magnetic disk drive, an optical disk drive, tape drive or the like. The storage device 160 is connected to the system bus 110 by a drive interface. The drives and the associated computer readable media provide nonvolatile storage of computer readable instructions, data structures, program modules and other data for the computing device 100. In one aspect, a hardware module that performs a particular function includes the software component stored in a tangible computer-readable medium in connection with the necessary hardware components, such as the CPU, bus, display, and so forth, to carry out the function. The basic components are known to those of skill in the art and appropriate variations are contemplated depending on the type of device, such as whether the device is a small, handheld computing device, a desktop computer, or a large computer server.
  • [0046]
    Although the exemplary environment described herein employs a hard disk, it should be appreciated by those skilled in the art that other types of computer readable media which can store data that is accessible by a computer, such as magnetic cassettes, flash memory cards, digital versatile disks, cartridges, random access memories (RAMs), or read only memory (ROM), may also be used in the exemplary operating environment.
  • [0047]
    To enable user interaction with the computing device 100, an input device 190 represents any number of input mechanisms, such as a microphone for speech, a touch-sensitive screen for gesture or graphical input, keyboard, mouse, motion input, speech and so forth. The device output 170 can also be one or more of a number of output mechanisms known to those of skill in the art. In some instances, multimodal systems enable a user to provide multiple types of input to communicate with the computing device 100. The communications interface 180 generally governs and manages the user input and system output. There is no restriction on the disclosed methods and devices operating on any particular hardware arrangement and therefore the basic features may easily be substituted for improved hardware or firmware arrangements as they are developed.
  • [0048]
    For clarity of explanation, the illustrative system embodiment is presented as comprising individual functional blocks (including functional blocks labeled as a “processor”). The functions these blocks represent may be provided through the use of either shared or dedicated hardware, including, but not limited to, hardware capable of executing software. For example the functions of one or more processors presented in FIG. 1 may be provided by a single shared processor or multiple processors. (Use of the term “processor” should not be construed to refer exclusively to hardware capable of executing software.) Illustrative embodiments may comprise microprocessor and/or digital signal processor (DSP) hardware, read-only memory (ROM) for storing software performing the operations discussed below, and random access memory (RAM) for storing results. Very large scale integration (VLSI) hardware embodiments, as well as custom VLSI circuitry in combination with a general purpose DSP circuit, may also be provided.
  • [0049]
    The logical operations of the various embodiments are implemented as: (1) a sequence of computer implemented steps, operations, or procedures running on a programmable circuit within a general use computer, (2) a sequence of computer implemented steps, operations, or procedures running on a specific-use programmable circuit; and/or (3) interconnected machine modules or program engines within the programmable circuits.
  • [0050]
    The computing device described above is particularly useful for generating and using mixes of media items based on a clustering analysis of similarity data derived from a large sample of users. Based on mathematical analysis to determine the incidence of correlation of two or more songs across multiple users' libraries, a statistical similarity can be determined. From the statistical similarity data, a clustering analysis can divide a large data set into smaller sets of related media items. By projecting the results of the analysis of the large data set into smaller data sets (media libraries of individual program participant's) high quality mixes of media items can be created.
  • [0051]
    This similarity data can be used for a variety of helpful functions including generating high quality playlists, determining suggested products for recommendation to a user for purchase, keeping a user's media device up-to-date with a fresh selection of songs, generating mixes from a user's media library among others. In one example, users may upload information relating to the contents of their entire media database to a server accessible to a wide audience of users. Based on how often two or more media items co-occur in the libraries of the entire audience of users, media items become correlated to each other and are noted as similar. Based on this similarity, the server can create a playlist to include one or more similar songs based on the selection of a seed track, or the server could recommend purchase of similar songs not in the user's library. These and other features of the system will be described in more detail below.
  • [0052]
    The clusters can be used in any way in which a collection of similar items can be used. In one example, clusters are useful in generating playlists based on the contents of a cluster. In another example, clusters are useful in generating mixes of media items. Mixes in this context are analogous to the concept of music mixes wherein a listener would create a CD or tape of a selection of music items that went well together. In this case, clusters provide the device with a method of automatically generating mixes of media items.
  • [0053]
    While many of the features of the described techniques and products will be described with reference to an online music store such as Apple Inc.'s ITUNES Store, it should be appreciated that the system does not require any media sales whatsoever to carry out the preferred embodiments of the system. Further, while some of the descriptions may refer only to one particular form of media, the principles described herein can be applied to any media, such as audio, video, still pictures, documents and other types of data files.
  • Server Uploading
  • [0054]
    FIG. 2 illustrates an example method embodiment for generating similarity data. The method includes receiving by a server, user data related to program participants' personal media inventory (202). The user data comprises identifying information regarding individual media items in a user's library usually in the form of metadata. The origin of the media items within a user's library is not relevant to the system. Even if the method described is employed by a music content store, the present embodiment does not require that a user's files originate from the music store since the server does not differentiate between file information based on source—all media items are treated the same. The method applies equally to physical media, such as CDs, DVDs, or HD-DVDs, as well as instances where the physical media is transcoded to digital format and entered into the user's media library. In a preferred embodiment, the server does not require that a user upload information regarding all media items in their library; information regarding only a limited number of media items or only partial information about the media items is also accepted by the server.
  • Identifying User Data
  • [0055]
    The method also includes identifying the user data as specific media items (203). This step may be accomplished in any one of a number of ways including examining the file name, metadata or through a detailed analysis of a portion of the file. In the example of a music track, the server can identify the music track through a variety of mechanisms. For example, a given track such as Led Zepplin's Stairway to Heaven having a length of approximately eight minutes can have a variety of data associated with the file including the file name or metadata identifying the artist, song name, genre, length of track, album name, track number, etc. A portion of this data can be sufficient for the server to identify the file as Led Zepplin's Stairway to Heaven from the album Led Zepplin IV. Alternatively, commercial song identification services can be used to identify the file. It is also possible to identify the track by sampling a portion of the music track or by recognizing a unique identifier from a purchasing store or by recognizing the songs digital fingerprint. For example, the server or service can recognize that a song was bought from the ITUNES Store and the file's metadata contains a unique identifier specific to the ITUNES Store. The server can identify the media item by that identifier. It should be appreciated that many other possible methods of recognizing media items are well known and all of which are encompassed by the disclosed method and arrangement.
  • Correlation
  • [0056]
    Each media item is associated with a unique identifier (204) and stored in a table (205). If the server has never encountered the song before, the server assigns a new identifier to that track. However, for most media items, an identifier will already be assigned and media need only be correlated with that identifier. Continuing with the example of Stairway to Heaven, once the media item has been identified as that track, the server can look up the unique identification data for that track. This process is repeated for each media item received by the server.
  • [0057]
    Although the server will already have an identifier to associate with most songs, in some cases, the media item will be completely new to the server, or at least recognized as such. In these cases, the server will issue a new identifier and assign it to that media item. For example, a new artist or garage band might not be recognized by the server. However, the server will issue a new identifier for that artist's track and when the server encounters the same track in a different user's library, it will assign the same identifier to that track as was previously issued.
  • [0058]
    In some cases, the server might not supply a new identifier to a media item that is new to the server. For example, a media item might be unique to a user's library and therefore providing an identifier would not be of any value since there would be no incidences of co-occurrence for that item. Accordingly, the system need not supply a unique identifier to all new items. Later, the system can supply a unique identifier for that track and supply similarity data for that item if and when other users also have that media item in their libraries.
  • Master Purchase Matrix
  • [0059]
    Each media item is stored in a master purchase matrix in step 205. An illustration of the master purchase matrix is shown in FIG. 4. The purchase matrix 402 is a table that contains columns 404 corresponding to each unique media item in the table. In the embodiment wherein the table is hosted by a music content seller, such as the ITUNES Store, the purchase matrix will also include all media items available for purchase from the ITUNES Store. The rows of the master purchase matrix 404 are associated with user unique user identification numbers 406. This identification code represents users by an anonymous identification code possibly known to the user, but not the system. Thus the master purchase matrix contains a record of every file received by the program server from every user that participates.
  • [0060]
    As the purchase matrix accommodates all program participants and all media in each program participant's library, the purchase matrix 402 is likely to be staggeringly large. For example, if a music store has approximately 2.25 million users and more than 5 million songs, a purchase matrix for such an online store would be a table of approximately 2.25 million rows and 5 million columns, or 1,125 billion individual cells. ITUNES Store has over 50 million registered users and contains over 10 million songs. A purchase matrix for such an online store could be a table of approximately 50 million rows and 10 million columns, or 500,000 billion individual cells.
  • [0061]
    For songs in a user's library, the cell in the matrix corresponding to that user and that song is marked. The marking may be done with a data type as simple as a Boolean, 1 for purchased 408 and 0 for not purchased 410. Other data types may be used when more data must be stored than a Boolean data type will allow. While the matrix can be very large, it is very sparsely populated because most users have relatively few songs (compared to 5 million) in their library. Each user row may only contain a handful of entries while each song column may contain tens, hundreds, thousands, or even millions of entries for more popular songs, but the vast majority of the purchase matrix is empty.
  • [0062]
    It should also be appreciated that the master purchase matrix or table will require updating as user libraries change, from adding new songs to their library, changing file metadata, or deleting files. The purchase matrix may be updated in real time or transactions may be queued up for insertion at a later date. One arrangement for queuing inserts transactions into the purchase matrix on a regular basis, such as every day at 11:30 p.m. Another queuing arrangement inserts transactions into the purchase matrix when some minimum number of transactions is ready for insertion.
  • [0063]
    Updates to the purchase table do not require repeating steps 203 and 204 for all media items. Media items that have already been identified and correlated to a program-based identification number do not require the identification and correlation step because it has already been performed when the user's data initially populated the table. In one embodiment, once the server has correlated the media item to its identification data, the server can send the identification data back to the user for storage in metadata. In future communications with the server, the server can receive updates to a media item with the associated identification data thus avoiding the need to redo step 204.
  • [0064]
    In some embodiments the purchase matrix may be part of a larger table or matrix including additional information regarding each transaction such as time and date of sale, what format the media is in, whether the media is high definition or standard definition, whether the transaction is a gift or whether it was paid for, which devices are authorized for playback, etc. The purchase matrix may be discrete, separate tables or may be wholly integrated into a larger table or matrix.
  • [0065]
    In still yet another embodiment, the master purchase matrix need not contain uploaded information about a user's media inventory. The master purchase matrix can be essentially comprised of purchase information. Purchase information can be only items purchased from the online store by users or can include additional information. In this embodiment the table can contain a list of all items purchased by each user. Still, the master purchase matrix would likely contain at least a limited set of data regarding the content of a user's media inventory.
  • Collabarative Filter Engine
  • [0066]
    The method illustrated in FIG. 2 next comprises running a calibration filter engine (shown as 308 in FIG. 14) on the master purchase matrix to generate similarity data (206) and store the similarity data in a master similarity table (207). Similarity data can be derived from a media item's incidence of co-occurrence with some other media item in the master table. When users have the same two media items, they are said to co-occur. Such co-occurrence in multiple users' libraries is an indication that the two media items are “similar.” The more often two media items co-occur in program participants' libraries, the more similar the media items are said to be.
  • [0067]
    The collaborative filter engine 308 calculates the similarity between individual media items in the master purchase table. In one embodiment the collaborative filter engine can be as simple as a program that tallies the number of times pairs of songs in the database co-occur in the program participants' libraries.
  • [0068]
    In some embodiments, the collaborative filter engine can calculate the similarity between two different media items by representing each item as a vector in a multidimensional vector space. The number of dimensions is equal to the number of users in the purchase matrix 306 (in FIG. 14). A correlation between items is developed by computing the distance between the vectors. A pure or modified trigonometric function (e.g. cosine or sine function) can be used to calculate the distance between vectors. If the two vectors have a small angle, they are considered to be similar and if the two vectors have a large angle, they are considered to be less similar. This process is carried out by comparing each song in the master purchase matrix to every other song in the master purchase matrix.
  • [0069]
    In some embodiments the collaborative filter engine can take into account more than the incidence of co-occurrence data. For instance, the incidence of overall occurrence can be factored into the similarity rating. For example, songs A and B may co-occur equally as often as songs A and C, but songs A and B can have a higher similarity score (that is, deemed more similar) than songs A and C if song B is more popular overall. One way of determining the popularity of the different songs is by measuring how often each song occurs in the overall data set.
  • [0070]
    Likewise there can be other inputs into the collaborative filter engine which affect similarity scores, such as usage inputs. For example, ratings given to the media items, play counts, and skip counts can all be inputs into the collaborative filter engine for determining similarity between media items.
  • [0071]
    FIGS. 3A and 3B illustrate the concept of representing each song as a vector. In FIG. 3A, a vector for Media ID 629 in the master table has been mapped out based on the songs existence in the libraries of users 2, 899, 1,343 and 2,000 Likewise, a vector has been mapped for Media ID 1,263,999 in the master table based on the song's existence in the libraries of users 2, 12, 444, 1,343. The angle between these two vectors is measured to be twenty degrees and calculates to a cosine value of approximately 0.939. The cosine value represents the correlation value between Media ID 629 and Media ID 1,263,999. FIG. 3B represents another correlation calculation to determine the correlation between Media ID 629 and Media ID 15. In this example only one user has both Media ID 629 and Media ID 15 in their libraries and therefore the cosine value between these two vectors is approximately 0.707. The lower score indicates that Media IDs 629 and 15 are not closely correlated to each other.
  • [0072]
    In another embodiment, other functions can be used to calculate the similarity scores between items. By way of example the similarity score can be calculated in a method that takes into account the percentage of an item's total co-occurrences that is made up by a particular item. For example item A co-occurs with items 1-5 and the item A's total co-occurrences is the sum of all the co-occurrences with items 1-5. For instance, if item A co-occurs with item 1 two times, item 2 two times, item 3 two times, item 4 ten times, and item 5 two times, then item A has 18 total co-occurrences. It can be valuable to learn what percentage of the total number of co-occurrences is due to a particular pair. For example, item 4 represents approximately 55% of all of item A's co-occurrences and therefore is more strongly correlated with item A than any of the other items which represent only about 11% each of the total of A's co-occurrences. This data can also be used to determine similarity data. Although two specific methods of calculating similarity data are illustrated above, it will be appreciated that any number of other methods of calculating the similarity between two items can be used.
  • [0073]
    The process above is computed between every item and every other item at some regular interval. As this process could take inordinate amounts of time, it is typically performed offline. It could be performed once a month, once a week, once a day, or as frequently as computing capability allows.
  • Similarity Table
  • [0074]
    The correlation data compiled by the collaborative filter engine is used to generate a master similarity table (207 of FIG. 2). FIG. 5 illustrates data stored in a master similarity table 502. For a given media item, there is an entry in the master similarity table identifying the media item by its program-based identification data 504. Each other media item considered to be similar or correlated to that media item is also listed in the table 508. In this way the table 502 maintains a list of all media that is similar to each media item in the master purchase matrix 402.
  • [0075]
    Items are included in the master similarity table if they have a sufficient score to be considered similar or correlated to the media item. For any given item, most of the other items are not similar at all. When the similarity score is sufficiently low, it is not included in the similarity table. The threshold for determining if a similarity score is sufficiently low or high may be dynamic or static. For example, if a similarity score ranges from −1 to 1, −1 being completely dissimilar and 1 being extremely similar, the threshold may be statically set to 0.5. The threshold may be dynamically set based on the number of occurrences of the song in the master purchase table, such as a threshold of 0.9 for songs occurring under 100 times, 0.7 for songs occurring under 5,000 times, 0.6 for songs occurring under 25,000 times, and 0.5 for songs occurring 25,000 times or more. The threshold may also be based on available storage or any other parameter.
  • [0076]
    In other words, the master similarity table 502 may incorporate all or some of the similarity scores. If the master similarity table must be constrained to fit a certain storage size, then only the best or strongest similarities are included and the rest are culled out. Similarly, if not enough program participants have an item in their media inventory, it could be excluded. One variation on this is to require a minimum number of occurrences in users' libraries before an item is eligible for inclusion into the master similarity table. For example, if two items have a co-occurrence of five or less, i.e. if five or less people have both of these items, the system does not compute the rest of the score.
  • [0077]
    In some embodiments the master similarity table can also store equivalency relationships. For example, media ID 17 in table 502 may correspond to “O Sole Mio” by Luciano Pavarotti, as performed in 1990. Multiple renditions of “O Sole Mio” by artists such as Enrico Caruso and Mario Lanza are available in the online store and are considered equivalent media 506. Other notable types of equivalent media include official album tracks, unplugged acoustic tracks, live concert tracks, cover tracks by other bands, and even foreign language tracks of the same song. A single song may exist in all the listed forms, and more. The determination of equivalency can be performed by hand or can be automatically performed based on the actual media content, metadata, and/or other available data. Equivalent media IDs for each are presented alongside the media ID 504 to identify that each of them is considered the same and that they are counted together.
  • [0078]
    The IDs in the master similarity table may be the same as those used in the purchase matrix 306 to index media or they may be based on an entirely different scheme. If the two do not align, a translator function or reference translation table can be provided to convert one ID to the other and vice versa. A common ID space shared between the two is likely to be the easiest and simplest to maintain without a performance penalty.
  • User Similarity Tables
  • [0079]
    Although the master similarity table contains the universe of similar media, a personal media library is almost certain to include a lesser subset of that universe of similar media. Thus, a constrained set, or an individual similarity table, is generated. The constraint is tailored to media availability in a given library. The constrained set may also include references to similar media not found in the media library. This can be done in order to target suggested media purchases to a user. Such targeted suggested purchases are likely to be more appealing to the user because other people with some common media tastes already have the suggested purchase in their library.
  • [0080]
    Once the master similarity table 502 is calculated and populated, individual similarity tables 602 are generated. FIG. 6 illustrates an individual similarity table. An individual similarity table 602 contains a media ID 604 for each song contained in the individual's media library and media correlated to that media ID. Correlated media 606 are shown in braces. Correlated media in an individual similarity table are a subset of the entire list of sorted similar items 508 located in the master similarity table of items. The subset is selected based on which media are present in the individual's media library. In other words, for each song in the individual's library, a list of similar media also contained in the individual's library is generated. Items 608 represent an alternate embodiment wherein the values are media IDs of songs that are not in the program participant's library, but are recommended for purchase.
  • [0081]
    Individual similarity tables can be generated by the server at the server and downloaded by client devices or individual similarity tables can be created by the client by downloading only similarity data for items in the user's media inventory. Regardless of where the similarity tables are created, the process is the same. Media items are looked up in the master similarity table by that item's program-based similarity identification data. Next the system can lookup the location of the similarity data for that media item in an index. Based on the results of the lookup operations, the system can retrieve the similar items. These items are further compared with those media items in a user's media library and only those items that are present in the user's media library are stored in the individual similarity table. In an additional embodiment, some highly similar tracks could also be recorded in the individual similarity table to be used for suggesting purchases to the user.
  • [0082]
    In some embodiments, at least one individual similarity table per library can be generated by the server. The client can send the library data to the server, which can identify the items in the library and generate the individual similarity table for that library by extracting the similarity data from the master similarity table pertaining to the items in that library. The resulting individual similarity table can then be downloaded to the client for use in generating playlists. The same individual similarity table can also be used by the client to create individual similarity tables for peripheral devices.
  • [0083]
    Using these individual similarity tables, users can benefit from data derived from the entire population of users. Users can use the similarity data to create playlists, receive suggestions for new media purchases, and a variety of other possibilities.
  • User Data Influencing Collaborative Filtering
  • [0084]
    In the embodiments wherein a playlist is generated based on similarity data, users can provide, or the system can require, feedback by skipping songs that they do not want to hear in that playlist or they can give positive ratings to media items that they enjoy hearing in the playlist. Such data can be uploaded from the client devices and received by the server. Feedback data can be stored in any number of different modes such as in the master purchasing index or in separate tables. Feedback can also be derived from online music stores such as the ITUNES Store. In this embodiment, user activity on the online store 304 (in FIG. 14) generates feedback data. User activity on the online store can include: purchasing media items, previewing media items, searching for a particular media item title, searching for a particular media item artist or searching for a particular media item genre. Feedback data can be used by the collaborative filter engine to modify correlation scores between items.
  • [0085]
    Feedback data from the population could potentially eliminate a correlation that otherwise exists. For example, the song Mrs. Robinson by Simon and Garfunkel may receive a strong correlation to the same song by the Lemonheads or Frank Sinatra, but users intending to generate a playlist based on the Simon and Garfunkel version might not like the different versions—perhaps due to the different genres of the cover songs. Over time, feedback data such as skipping the song or removing the item from the playlist could be used to eliminate one or both cover versions from the list of correlated items.
  • [0086]
    In some embodiments, user feedback can also be used to modify similarity data directly on the client device without sending the feedback to the server first for use by the collaborative filter engine. In these embodiments, user feedback data can directly effect the relationships stored in the individual similarity table on the client device without first synching with the server. Alternatively, the similarity table can remain the same, but constraints can be generated based on user feedback data. For example, if a user skips a similar item often, the client itself can either no longer consider that item as similar by removing the item from the individual similarity table or can prevent the addition of the frequently skipped item from inclusion in the playlist by other means, such as a constraint in the playlist algorithm.
  • Updating Similiarity Table
  • [0087]
    Over time, new media items can be added to the online store, and by extension, the master purchase table, as artists create new music and new artists emerge. Accordingly, a correlation between existing items in the online store 304 and newly added items in the online store may form. For example, if a user has Song A in his/her media library and subsequently purchases a newly added Song B that was recently introduced on the online store 304, then a new correlation between Song A and newly added Song B may form. When a user purchases newly added Song B, the transaction is recorded in the purchase matrix 306 and an updated purchase matrix 306 is generated. Alternatively, Song B may have been added to a user's library by copying the track from a CD. In such a circumstance, the result would be the same. As long as both Song A and Song B are in the user's media library, this co-occurrence is recorded in the updated master purchase matrix.
  • [0088]
    The collaborative filter engine 308 uses information from the updated purchase matrix 306 to generate an updated master similarity table of items 316 as described above. Thereafter an updated individual similarity table 322 can be generated for each user by extracting only those items in the updated master similarity table 316 which have been changed. Alternatively, an entirely new individual similarity table can be created based on the new data.
  • Clusters
  • [0089]
    The similarity data discussed above can also be utilized in a clustering analysis to identify groups of media items. At a conceptual level, the clustering analysis graphs items based on the similarity vectors and selects items that are plotted relatively closely together in the graph. For example, FIG. 7 illustrates a graph of several media items 420-428 as similarity vectors representing media items and every media item similar to that item, wherein the vectors are grouped together to form three clusters of similarity vectors. A cluster analysis groups media item 420 in its own cluster 430, media items 421-426 in cluster 432 and media items 427-428 in cluster 434.
  • [0090]
    In a music context, the clusters 430, 432, and 434 could represent media items belonging to separate genres, sub-genres, or just groups of media items that are more related to each other than other media items.
  • [0091]
    FIG. 8 illustrates child clusters of cluster 432. FIG. 8 shows the same similarity vector space as illustrated in FIG. 7, however, media items 421-426, which were grouped into cluster 432 in FIG. 7, are subdivided into child clusters 435, 436, and 437. Child clusters 435-437 can be generated by reapplying the cluster analysis on the smaller group of items contained in cluster 432.
  • [0092]
    FIG. 9 illustrates the clusters computed in FIGS. 7-8 in a tree diagram. At the top level, CLUSTER 0, contains all of the media items represented in FIGS. 7-8. The second level having clusters 432, 430, 434 correspond to the same clusters in FIG. 7. The child clusters at the bottom of the diagram 435, 436, 437 correspond to the same clusters in FIG. 8. More specifically, performing a clustering analysis on CLUSTER 0 arrives at the clusters shown in FIG. 7 and illustrated by the middle nodes in FIG. 9. Likewise, performing a cluster analysis on parent cluster 432 results in child clusters 435, 436, 437 shown in FIGS. 8 and 9.
  • [0093]
    There are many known mathematical forms of clustering analysis. In the present technology mechanisms of hierarchical clustering are employed. Hierarchical clustering creates a hierarchy of clusters that may be represented in a tree structure, as discussed above. These algorithms can be either agglomerative (“bottom-up”) or divisive (“top-down”). Agglomerative algorithms begin with each element as a separate cluster and merge them into successively larger clusters. Divisive algorithms begin with the whole set and proceed to divide it into successively smaller clusters. A divisive clustering analysis was illustrated in FIGS. 7-9. It will be appreciated that many techniques for clustering analysis exist and can be applied in the present technology.
  • [0094]
    In some embodiments the type of clustering analysis applied is art as a k-means analysis. The k-means algorithm assigns each point to the cluster whose center is nearest. The center is the average of all the points in the cluster—that is, its coordinates are the arithmetic mean for each dimension separately over all the points in the cluster. For example: if the data set has three dimensions and the cluster has two points: X=(x1, x2, x3) and Y=(y1, y2, y3). Then the center Z becomes Z=(z1, z2, z3), where z1=(x1+y1)/2 and z2=(x2+y2)/2 and z3=(x3+y3)/2.
  • [0095]
    A k-means analysis performs the following steps: (1) Choose the number of clusters, k; (2) Randomly generate k clusters and determine the cluster centers, or directly generate k random points as cluster centers; (3) Assign each point to the nearest cluster center; (4) Re-compute the new cluster centers; (5) Repeat the two previous steps until the lowest level clusters have less than a determined number of media items within each lowest-level cluster. An example implementation of a K-means analysis is illustrated in FIGS. 10-12.
  • [0096]
    In some embodiments, the database of media items is divided into canopies (102) based on any reasonable criteria. While not necessary to achieve clusters of highly similar media items, this step can be useful in a variety of ways. First, in the case of large datasets, a simple division of the database can reduce the computation resources required to compute clusters for the top level of the database. Instead of performing a statistical clustering analysis on the entire database of millions or billions of media items, the statistical clustering analysis can be performed on smaller subsets of the whole.
  • [0097]
    Additionally, dividing the database into canopies can also enhance the quality of the resulting clusters. For example, the canopies can be generated based on editorial divisions. In FIG. 13, canopies 450, 451, and 453 are examples of divisions of an entire database of music items into non-overlapping groupings of non-similar genres of music media items. In such instances, the editorial decision that varieties of rock 450, rap 451 and opera 453 music media items are never similar to each other has been made and the database of media items has initially been split into these canopies.
  • [0098]
    In some embodiments the canopies can be divided into sub-canopies based on editorial decisions. However, in preferred embodiments, the canopies should only be divided so that the canopies or sub-canopies separate items that should never be grouped together because items in two different canopies cannot later be grouped in the same cluster.
  • [0099]
    Next, the chosen clustering algorithm can be performed on each canopy/sub-canopy to further subdivide the database into clusters. As illustrated in FIG. 13, the rock canopy 450 has been divided into two sub-groupings of similar media items, pop/alt/classic rock 454, and hard/metal rock 455. Likewise the rap canopy 451 has been further divided into east coast rap 459 and west coast rap 460 clusters. It should be noted that while FIG. 13 illustrates the clusters being sub-divided by genre and sub-genre, this is for purposes of illustration only. The statistical clustering algorithm will sub-divide clusters according to the statistical analysis of similarity data, which is based on occurrences of media items in the libraries of program participants, and will likely produce clusters having media items in a variety of compatible genres.
  • [0100]
    The clustering analysis can be performed as follows: The items in a given canopy or sub-canopy are randomly divided in to X (X being an arbitrary number greater than 1) number of clusters (1013). For each cluster a centroid is calculated.
  • [0101]
    FIG. 12 illustrates one method of calculating a centroid. 1050 is a collection of similarity vectors for the media items in a given cluster. The similarity vector lists media items that are the most similar to the media item in order of descending similarity. For example 1050 illustrates a vector for items 1, 2 and 25 which are grouped in a cluster. Item 1 has media item 1 as most similar to it and item 25 as second most similar and so on. Each item in the cluster is then given a value based on how similar the item is to the other items in the cluster. This is accomplished by assigning weighted values (weighting more similar items greater than less similar items) to each item and adding up the values. For example, 1052 illustrates a table having each item found in the similarity vectors of the media items presented in the cluster and their associated weighted values. Notice that not every media item in the table is actually in the cluster.
  • [0102]
    Item 1 was most similar to Item 1 and thus is given a value of 100 and item 25 was second most similar and is given a value of 99. The scores for each item are totaled to result in each item's score. For this cluster Item 1 receives a score of 199. Item 25 is highly ranked in each list and accordingly receives a higher score of 298. The scoring is calculated for each item. The final scores can be sorted and used to create a similarity vector for the entire cluster 1054. This process is repeated for each of the X number of clusters to result in X centroids.
  • [0103]
    Many different scoring mechanisms are conceivable and are considered within the present technology. The illustrative method described above is just one method of calculating a centroid for the given cluster. In the present example the centroid is effectively an average of all the media item's similarity vectors of those media items within the cluster.
  • [0104]
    Once the centroids have been calculated, the randomly created clusters can be discarded. Their purpose was only to establish centroids. Next, the similairity vector for each media item within the canopy is compared to each centroid (also a vector) (1015) and each media item is assigned to the cluster containing the nearest centroid (1016) to create X clusters.
  • [0105]
    The process is run iteratively to generate the highest quality cluster by computing the centroid for the existing cluster (1014), and then comparing each items similarity vector to each new centroid (1015) and clustering items to the nearest centroid (1016). The process ends (1026) once either the maximum number of iterations has been completed (1022) or media items no longer migrate from one cluster to another (1024). In this way, the iterative process slowly refines the cluster analysis by refining the centroid and migrating songs into the appropriate cluster.
  • [0106]
    The entire process can be re-run on any given cluster. For example, the process can be re-run on any of the X clusters created above to create sub-clusters as illustrated in FIG. 11. Anyone of the X clusters can be randomly divided into X sub-clusters (1028) (again, X can be any arbitrary number greater than 1). The centroid for each sub-cluster is calculated (1030), each media item's similarity vector is compared to the vector for each centroid (1034) and grouped into clusters (1036). A new centroid is calculated (1030), and each media item is compared to that centroid (1034) and grouped into clusters (1036). The process repeats until the max number of iterations has been completed (1046) or no media items change clusters (1048) and the process ends (1050).
  • [0107]
    The process can continue to repeat until the population of media items has been reduced to a desired number of clusters or clusters of a desired size.
  • [0108]
    In an alternate variation on the clustering method described above, instead of clustering the media items individually, the items can be clustered by artist. Just as described above, the media items will be divided based on editorial decisions which eliminates the possibility of non-similar songs by the same artist from being group together.
  • [0109]
    For each artist represented in a canopy, a vector can be created in the same way the centroids are created and illustrated in FIG. 12. Each song by an artist in a canopy is grouped together and an artist vector is generated. The vector contains each song from the artist that is the canopy. Having computed vectors for each artist, the clustering algorithm can use the artist vectors in place of media items and organize the artists into clusters. The iterative clustering process is complete once the process has repeated a selected number of times or only a chosen number of artists remain in each cluster, for example 3 artists. Having completed the clustering process, the artist vectors can be returned into media items by identifying each song in the artists' vector.
  • [0110]
    The process described above can be used to generate clusters of every media item in a database on a server. Once the entire database has been divided into a complete tree of server clusters, the tree of server clusters can be used to generate participant clusters. Participant clusters are specific to an individual user's library and contain only media items contained in the participant's media library.
  • [0111]
    A program participant's media library can be mapped to the server clusters by finding each media item in a program participant's media library in a lowest-level server cluster in the tree and working up the tree of clusters until a participant cluster having at least a determined number of items is accumulated. This method is illustrated in FIG. 14. FIG. 15 illustrates a program participant's library mapped to a tree of server clusters labeled according to hierarchy. The tree represents the same relationship as the entire server cluster tree structure, but the clusters are shown illustrating the number of a program participant's media items that are in any one of those server clusters (clusters without any numbers within them do not contain any of the program participant's media items).
  • [0112]
    A program participant's media items are mapped to the lowest level server clusters (470). See, for example FIG. 15, server cluster 1.1.1.1 contains 3 of the program participant's items. Table 1, below identifies the number of a program participant's media items in the lowest-level child clusters.
  • [0000]
    TABLE 1
    Number of Program
    Participant's Media
    Cluster Id Items
    1.1.1.1 3
    1.1.1.2 2
    1.1.1.3 1
    1.1.3.1 5
    1.1.3.2 3
    1.1.3.3 4
    1.1.3.4 5
    1.2.2.1 10
    1.2.2.2 12
    1.3.1 1
    1.3.2 1
    1.4.1.1.1 5
    1.4.1.1.2 5
    1.4.1.1.3 5
    1.4.3.1 2
    1.4.3.2 1
  • [0113]
    Returning to FIG. 14, the process determines if the cluster contains more than a determined number of items. In some embodiments the determined number of items can be 10 or greater, 20 or greater or 50 or greater. In some embodiments the size of the cluster can also be limited to a determined maximum. For the purposes of the present example the determined number of items is 20 or greater. Thus at 471 the process determines that server cluster 1.1.1.1 contains less than 20 items and proceeds to 472 and adds media items from other participant clusters into a common parent. Since cluster 1.1.1.1 shares a common parent with clusters 1.1.1.2 and 1.1.1.3 containing 2 and 1 of the program participant's media items respectively, these media items are merged into one larger participant cluster. Again the process checks the number of media items in the participant cluster and since only 6 items are present in the cluster the process repeats (472). This time, server clusters 1.1.1, 1.1.2 and 1.1.3 all share a common parent and by merging these clusters a participant cluster of 28 media items results. Now when the process checks for a determined number of items (473), the participant cluster exceeds 19 and the process ends 474. The cluster is designated as 475 in FIG. 15.
  • [0114]
    As discussed above, this process of grouping similar items together by merging clusters higher in the tree is called agglomeration.
  • [0115]
    FIG. 15 illustrates several other potential outcomes of the process illustrated in FIG. 14. For example participant cluster 476 only required one iteration to have agglomerated a sufficient number of items. Cluster 477 illustrates that the lowest level of the tree diagram for every branch will not necessarily be the same. Cluster 477 began at a 5th order server cluster and merged all 5th and 4th order server clusters into the same parent to result in participant cluster 477.
  • [0116]
    478 illustrates a failed cluster. In order to agglomerate enough items to achieve the desired size, 478 would need to encompass the top order server cluster. In such an instance, it can be concluded that these items to do exhibit sufficient similarity to be combined into a cluster. The cluster could fail because a programming decision has been made to disallow agglomeration beyond a certain level of the tree, or the cluster could fail for now having too many items.
  • System Overview
  • [0117]
    FIG. 16 illustrates an example of one such system 300 that can carry out the embodiments described above. Purchase Matrix 306 stores data regarding all media items from each user's library. The collaborative filter engine 308 uses the data in the purchase matrix 306 to determine the incidence of co-occurrence of each media item relative to each of the others. The collaborative filter engine 308 can optionally use feedback data stored in objects 324 and 326 to adjust similarity scores according to user preference across the population of program participants. The output of the collaborative filter engine is stored in the master similarity table by storing all media items that are deemed similar 320 to a given media item 318. The similar media items 320 can be sorted to be presented in order from the most correlated items to the least correlated.
  • [0118]
    Likewise, the collaborative filter engine can also process its inputs to divide the database in a tree of server clusters 332. Alternatively, the tree of server clusters 332 can be generated from the data output by the collaborative filter engine. Either data set is appropriate for a clustering analysis.
  • [0119]
    The online management tool 304 serves as the interface for the client side 302 and the server side 301. In a preferred embodiment, the online management tool 304 can be associated with an online store. In one embodiment, the online store may generate data 310 which can also be used by the collaborative filter engine 308 in generating similarity data. Regardless, the client 302 serves to both upload information regarding the media items stored on a client 302 and to download similarity data and clustering data from the server 301. Similarity data specific to the client may be downloaded through the online management tool 304 directly from the master similarity table 316 to generate an individual similarity table 322 on the client device 302 or the individual similarity table may be generated by the server 301 and downloaded to the client device 302. The client device can also communicate feedback data through the online management tool 304 to the collaborative filter 308 to aid in generating better similarity data.
  • [0120]
    Similarly, the client device 302 can download clustering information from the server 301. In some embodiments, the server 301 calculates participant clusters for a program participant and sends information identifying the participant clusters and the media items belonging to the participant clusters 330. In some embodiments, the server could send information detailing the presence of each of the program participant's media items in each server cluster and let the client agglomerate the participant clusters.
  • Client
  • [0121]
    As is illustrated in FIG. 16, the client device 302 can take any one of a variety of forms. Client devices 302 include devices such as portable digital audio players, portable digital video players, smartphones, desktop and laptop computers, television set top boxes, and any other device capable of playing media. Client devices can include offline portable devices which sync to the online store 304 through a desktop computer or other intermediate device, such as an IPOD digital audio player syncing through a cable connection with a desktop computer, an IPOD TOUCH digital audio player syncing wireles sly through a wireless router, or a smartphone syncing wirelessly through a cellular connection. As is apparent from the discussion above, the client device contains an inventory of a variety of media items. If the user of the client device wishes to benefit from the use of similarity data, the user accesses the service through an online interface. Alternatively, the client device can access the similarity data service through a client based application such as the ITUNES desktop application which communicates with a remote server via the Internet.
  • [0122]
    The online store 304 can be a single server or a cluster of servers which allow client devices to purchase digital media. The online store may also grant access to other ancillary media management features.
  • [0123]
    FIG. 17 illustrates a method embodiment for downloading similarity data to the client. If this is the client's first time accessing the collaborative program 842, then the client must choose to participate by opting into the program 844. The client is given a private user identification code at step 846. While the server will use this code to associate with the client's library, only the client contains additional identifying information. The server knows no personal or identifying information regarding either the client device or the user (program participant). Once the private identification code has been assigned, the client device can upload information regarding its media library 848. In some embodiments, a user may select which media items to share with the server.
  • [0124]
    If the outcome of step 842 indicates that the user is already a participant in the collaborative program, the client can log into the server using its private identification code in step 850. In step 852, the server may request information from the client regarding changes to the media library or alternatively the client may share this information without first being requested. If changes to the library have occurred, the client uploads information regarding the changes 854. Changes can take the form of edited metadata, new items, deleted items, etc. Once up-to-date information regarding the client's media library has been sent to the server, the client can request new similarity and/or cluster data (collectively “participant data”) from the server in step 856. Alternatively, the server may initiate the transmission of new or updated participant data.
  • [0125]
    In some embodiments the client can skip the above steps and simply request updated results from the master similarity table without the need to identify the library. For example, the client can request an update for a particular item or list of items without identifying itself by identification number and without notifying the server of changes to the client's media library inventory.
  • [0126]
    Similarity data and/or cluster data is downloaded in step 858. Similarity data can be downloaded in the form of a table previously created by the server or the information can be downloaded and formatted into an individual similarity data table in real time. Cluster data can be downloaded as a list of clusters and media items belonging to the cluster. The data can be provided in the same or separate data structures or transmissions.
  • [0127]
    Playlists are generated based on the individual similarity tables. The playlists can be generated at the client side 302, in the online store 304, or in the collaborative filter engine 308. Playlists can be generated based on the statistical similarity to one song or multiple songs. When a client device employs playlists based on individual similarity tables, certain songs may not be what the user expected or may not be pleasing to the user. When a user often skips a song that is put in the playlist based on statistical similarity, the system “penalizes” the song, reduces its similarity value, includes it in playlists less frequently, and can eventually discard it entirely from playlisting. This is effectuated by reporting when a user skips a song. That act is recorded on the client device and transmitted as feedback to the online store. The online store sends this feedback to the collaborative filter engine which stores it in a negative correlations table 326. The negative correlations table stores negative indications of similarity. In this manner each user does not need to individually skip a song many times before it is rejected from the similarity table. The cumulative negative feedback will filter out unpopular songs for the group of users based on group behavior. As such, these playlists act as living organisms which evolve to suit the changing media tastes of the population of users.
  • [0128]
    For example, if many users always skip a particular song in a playlist or if many users delete the song from their playlists, that information is recorded and assigned a weight to influence the generation of the master similarity table. The weight may be based on the number of negative correlations reported or on other aspects of user interaction. Conversely, positive user interactions with a particular song provide positive correlations. For example, if many users add the same song to an existing playlist or if many users turn the volume up during a particular song, those positive correlations are reported through the online interface to the server and are stored in a positive correlations table.
  • [0129]
    In addition to interacting with the server, the client may also interact with other clients or peripheral devices. In one preferred embodiment, the client can be a personal computer which interacts with a portable music player such as an IPOD portable music player. In such an instance the client device can also prepare individual similarity tables for use on the peripheral device. Just as the server can prepare an individual similarity table for a client containing information for only those media items in the client's library, the client can prepare an individual similarity table for the peripheral device containing information for only those media items in the peripheral device's library. The client can load the individual similarity table onto the peripheral device just as it would any other data item.
  • Peripheral Device
  • [0130]
    The peripheral device is meant to work with the client just as the client works with the server. For example, the peripheral device can notify the client of changes to its contents and request similarity data for those contents. In a more preferred embodiment, the peripheral device can be managed by the client just as a personal computer running the ITUNES desktop application manages an IPOD portable media device. In this embodiment, the client usually does not need to be updated with changes to the peripheral device's contents because the client already knows the changes by being the tool that effected those changes. In this embodiment the client can also keep the peripheral device's similarity table up-to-date based on the available data in the client's individual similarity table. Further, the peripheral device can communicate feedback data to the client for later transmission to the server.
  • [0131]
    Again, cluster data can be sent to the peripheral device in a similar manner as to how the similarity data is updated. The client can send a cluster ID along with only media items present on the peripheral device.
  • [0132]
    In some embodiments, the peripheral device can interact directly with the server as a client device itself. As more and more portable media players have capabilities for accessing the internet, the portable device could interact directly with the server to download the most up-to-date similarity data or to update the master purchase matrix. In such an embodiment the peripheral device could have all the capabilities of the client.
  • Generating Playlists
  • [0133]
    As has been discussed, one use of the similarity data generated by the methods described herein would be to use similarity data to generate playlists. While generating playlists can be as simple as playing all similar items in the individual similarity table, the best results will likely be obtained through the use of a playlist generation module. FIG. 18 demonstrates a method embodiment for the generation of a playlist. At step 802 one or more seed tracks are selected and are added to the playlist at step 804. In step 806 the similarity data for the seed track(s) is looked up in an individual similarity table and the similar tracks are added to a candidate list in step 808. Candidates are a list of potential tracks to be added to the playlist. Whether a candidate is added to the playlist is determined by evaluating the candidate against a series of constraints.
  • [0134]
    Constraints serve as limits on whether a song can be added to a playlist. In most cases the constraints will serve to enhance the quality of the playlist. One constraint may require certain spacing in the playlist. For example, two songs by Radiohead could not play back-to-back if the artist spacing constraint were active. In a similar example, constraints may prevent songs from the same album or song title from occurring within a given number of songs of each other in a playlist. Another constraint would prevent songs from non-compatible genres from playing in the same playlist. Still yet another could be a “jitter” function. Jitter can randomly prevent a song that would otherwise be acceptable from being added to the playlist. Jitter provides randomness to a playlist to prevent the same playlist from being generated every time based the selection of the same seed track. Still yet another constraint could be a skip count constraint wherein any song that has been skipped more than a given number of times would fail the constraint and not be included in the playlist.
  • [0135]
    It should be appreciated that the constraints are dependent on certain parameters or variables that can be easily varied. For example, the required number of intervening songs between two songs from the same album or artist can be varied. Similarly, the number of skip counts needed to exclude a track or the members of compatible genres can be varied. In one embodiment, such variable settings can be configured by the user. In another embodiment, such settings are chosen by and issue from the server and may be adjusted by the server whenever the client connects to download updated similarity data.
  • [0136]
    It should also be appreciated that positive constraints are also contemplated. Songs with high play rates or high ratings can be moved up in the playlist or can be immune to the jitter function or effects of some other negative constraint. Many variations are possible by adding more restraints, removing constraints, making constraints more or less strict, each of which is contemplated by the method described herein.
  • [0137]
    In step 810 the next candidate to be considered is selected and its metadata is retrieved in step 812. The metadata provides information about the candidate that will be needed in evaluating the constraints. In steps 814, 815, and 816 the track is evaluated by the constraints. While only three constraints are represented in FIG. 18, it is contemplated that there can be any number of possible constraints. If the song passes one constraint, it is passed to the next constraint until all constraints are passed in step 818 and the track is added to the playlist. Thereafter the method proceeds to step 820 and fetches similarity data for the new song in the playlist. That similarity data can be appended to the candidate list for consideration for inclusion in the playlist.
  • [0138]
    If a track fails a constraint the method skips to step 830 and returns to step 810 if there are additional candidates in the list to evaluate. However, if there are no additional candidates in the list, meaning that all candidates either failed the constraints or are in the playlist, step 832 checks the constraints to determine if they are set to their default value. If they aren't, the constraints are relaxed at step 828 in the hope that relaxing one or more of the constraints will allow a previously rejected candidate to be added to the playlist upon reconsideration at step 826. If all candidates once again fail the constraints, step 832 determines that the constraints were already relaxed and ends the process.
  • [0139]
    Relaxing constraints can take any one of a variety of forms. In some cases only one constraint becomes less strict. In other cases multiple constraints are relaxed. For example, if the first constraint requires that two songs from the same artist do not play within 4 songs of each other, the constraint can be relaxed to only require one intervening song or the constraint can be eliminated entirely. In some embodiments, not all constraints can be relaxed. An example is the genre constraint. Certain genres may never be compatible and thus the genre constraint would not be relaxed to include additional genres.
  • [0140]
    Returning to step 826, in the instance where the candidate track failed the constraints and no additional candidates are available, the constraints are relaxed. If, due to the relaxed constraints, a candidate passes all constraints and is added to the playlist, the constraints are returned to normal at steps 822 and 824. Step 822 checks the constraints to determine if they are relaxed or are at their default settings. If they are at their default settings, the method returns to step 810 and considers the next track in the list. If, however, step 822 determines the constraints are relaxed, the method proceeds to step 824 which returns the constraints to normal and then returns to step 810 to consider the next candidate in the list. This process continues until no candidate passes the relaxed constraints and the routine ends at step 834.
  • [0141]
    Songs need not be added to a playlist in order of the most similar to least similar; songs can be added in any order using a randomization algorithm or other logic.
  • [0142]
    Generated playlists can be static or dynamic. Dynamic playlists can be updated over time as the similarity data evolves and as new items are added to a user's library. Alternatively, when a playlist is generated that a user enjoys and wants to retain without change, the user can save the playlist as a separate, static playlist which is no longer influenced by the similarity data. In this manner, when users feel that a particular generated playlist is “perfect”, it can be preserved unsullied from the changing tastes of the masses reflected in the similarity table through the server.
  • [0000]
    Mixes from Participant Clusters
  • [0143]
    The participant clusters are useful in generating mixes of media items. As an exemplary embodiment, the mixes can be music media item mixes.
  • [0144]
    As discussed above, a client device downloads information regarding the clusters contained within the program participant's media library. While any number of clusters can exist, in some embodiments it is desired to only keep track of a limited number of clusters, for example twelve. Since each of the clusters includes media items that do not overlap with any other cluster, each cluster will represent different collections of similar media items. The result is that each of the twelve clusters can represent different types of music. A program participant might be in the mood to listen to their classical media items, or their popular media items and the clusters can create mixes from these collections.
  • [0145]
    FIG. 19 illustrates the basic method for generating the mixes. Since the mixes are intended to be a representation of media items within the program participant's media library that are similar, an average similarity or affinity for the cluster is calculated 865. This can be done in a number of different ways including taking an average of all similarity vectors for the media items represented in a cluster. The average vector will represent the average affinity of the media items for the cluster. An alternative method is to choose a media item for which the similarity scores of the other media items with respect to the chosen media items are strongly similar and use that media item as a representative media item for the cluster.
  • [0146]
    The process can then select several songs (without any user involvement) from the cluster. Ideally, these items will have an approximately average similarity score to the cluster average so that the seed songs are representative of the cluster. In some embodiments it is desired to avoid selecting the most strongly correlated or the most weakly correlated items to the cluster. But, in some embodiments, the most strongly correlated items can be selected. And in some embodiments the selection of the several media items can be a random selection of media items from within the cluster.
  • [0147]
    The selected media items 867 are input into a playlist algorithm 869 such as the one described above. Each of the selected items constitutes a seed track for the playlist algorithm. Since the playlist module will lookup similar items 808, 820 for all items in the playlist to use as candidates for inclusion into the playlist, the playlist may eventually include media items outside the cluster. Once the playlist is created, it is played 871 until the program participant discontinues playback. As illustrated in FIG. 19, once the playlist has reached its end, the process continues so that the mix is endless. The result will be that media items that are generally similar to the other items in the selected cluster will continuously play.
  • Uses of Similarity Data for Syncing Devices
  • [0148]
    Just as playlists can be generated, similarity data can also be used to keep a peripheral device such as a portable music player updated with new or different music from the user's media library. Many portable media players suffer from limited capacity and thus not all of a user's media library can be stored on a portable device such as an IPOD portable media player. In such instances users often must select a portion of their media items to be stored on their portable device. Some user's may consider this process burdensome and may find that they get tired of the media items that are currently stored on their portable device.
  • [0149]
    One solution to the problem identified above is to use similarity data to keep the portable device updated with new media items. In this embodiment a user selects a certain number of seed tracks that represent the type of music that they would like on their portable device. Just as a playlist can be generated from seed tracks, so too can data items from a user's library be chosen for inclusion on a portable media player. As similarity data changes, and as new songs are added to a user's library, similarity data can be used to keep the portable device up-to-date with new tracks.
  • Purchase Recommendations
  • [0150]
    Similarity data can also be used to recommend new items for purchase by a user. Items that occur in the master similarity table, but are not present in the user's library can be recommended for purchase by the system. In one embodiment the online store can make purchase recommendations based on the user's library or the selection of a seed track. The online store can also recommend items for purchase using recent purchases as the seed track. In another embodiment the playlist module can recommend songs for purchase by informing the user of where the song would have been included in the playlist if it were part of the user's library. Purchase recommendations can occur at any level of the system from the server to the client to the portable media player.
  • [0151]
    Similarly, items in a sever cluster that are not owned by a program participant can be suggested for purchase to the program participant.
  • [0152]
    In some embodiments, the item to be recommended can be a new item that does not yet have similarity data associated with it. In such a scenario, a similarity score can be provided by an administrator of the online store so that the new item will be recommended for purchase by users. The provided similarity score can remain until similarity data based on co-occurrence or other embodiments is available.
  • [0153]
    Similarly, movie rentals or purchases could also be recommended. While most of the embodiments have been described with respect to songs, media items can also be any file including videos or movies. In this embodiment, a user can request similar movies based on the selection of a seed item. Alternatively, the system can recommend movies based on previous rentals. It is even contemplated that similarity data can be used to automatically download new movies to a client or portable media player for the user's viewing at a later time. This embodiment could be part of a subscribed service or the user could choose to not watch the movie and not-incur a rental fee.
  • [0154]
    FIG. 20 illustrates one method embodiment in which similarity data can be used for recommending media items for purchase. In step 1002, a seed track is selected. Purchase recommendations are based off of a seed track, just as playlists are generated based on a seed track. However, in at least one embodiment, purchase recommendations can be based on content most similar to a user's entire library.
  • [0155]
    Once a seed track is selected, the illustrated method next looks up similarity data for that seed track from the individual similarity table in step 1004. In this embodiment, it will be appreciated that the individual similarity table will also contain media items that are not in the user's library, but are related by similarity data to items that are present in the user's media library. In another embodiment, the purchase recommendation can come directly from the server, rather than the locally stored similarity table. Similarity lists from the master similarity table can provide similarity data and the server or client can subtract out the tracks identified from the user's library. This embodiment has the benefit of generating the smallest possible individual similarity tables, since purchase recommendations do not need to be included, and all users share the master table for recommendation purposes.
  • [0156]
    To prevent the items that are already in a user's library from being recommended for purchase, step 1006 filters items that are already in the user's media library from the similarity data retrieved in step 1004. In some cases media items have already been purchased by a user, but those items are not in the user's media library. This could be for any number of reasons such as: the user did not like the song and deleted it, or the user has stored the media item in another library, or the user may have purchased the media item but has not downloaded it yet. In at least these scenarios it would not be desirable to recommend these media items for purchase, so they too are filtered out in step 1008. Step 1010 displays the recommended items for purchase.
  • [0157]
    Of course, the method can repeat for every song that is currently selected by the user. In some embodiments, the user need not affirmatively seek purchase recommendations. In such embodiments, purchase recommendations can be displayed within a media organizer or media store whenever a media item is selected for playback. As the media item is played, a graphical user interface can display other recommended media items, or groups of media items, such as albums, for purchase. In a preferred embodiment, the media items can be purchased directly by selecting, with an input device, the recommendation itself, or at least the recommendation can hyperlink to an online store so that the media item can be purchased therefrom.
  • [0158]
    In another embodiment, purchase recommendations can also be shown using this methodology where a selection is not available. For example, in a View An Artist page which would display top songs by a designated artist that listeners have also bought, the system can display recommendations that are filtered to eliminate recommendations of items already in the user's library.
  • Server>Client>Ipod
  • [0159]
    In one particularly preferred embodiment, the methods that are useful at one location are equally applicable at each location in the system. FIG. 21 illustrates this concept. An online music store 902 is shown containing a similarity table 908 and a playlist module 914. The similarity table 908 contains all of the relationships for each item in the server's media inventory. The playlist module 914 operates based on the data in the similarity table just as it would on any other client or portable media player in the system. Likewise, the client 904 is shown containing a smaller media inventory than the server contains and similarity data for each item in the client's media inventory is stored in its similarity table 910. Playlist module 916 operates based on the data in the local similarity table 910 just as it would on the portable media player 906 or music store 902. Finally, portable media player 906 is shown containing a smaller media inventory than that of the client 904. The similarity table 912 stores information relevant to items in the portable media player's 906 media inventory and the playlist module 918 operates based on that data.
  • [0160]
    FIG. 21 illustrates the concept that the client can behave like a server to a group of portable media players or a larger client can behave as a server to smaller clients; all of which is contemplated herein.
  • [0161]
    In most embodiments, the server will be the only location maintaining the master purchase matrix and master similarity table, but the generation of local similarity tables can occur on any device based on the data derived from the server. Another function that will likely be carried out by the server is updating. Not only will the server be the source of similarity data updates in most embodiments, but the server can also update the playlist module or software components of the system.
  • [0162]
    It will be appreciated that not all processes described herein must occur in real time. In many cases it will be desirable to save processing resources and carry out certain tasks offline. For example, the collaboration filter engine may run only at scheduled times to generate the master similarity data table. It is further conceived that the need for offline processing may be relieved as technology advances. In such instances, those processes that are now preferred to be carried out offline could be carried out in real time.
  • [0163]
    In another embodiment, although each device is capable of carrying out certain functions locally, it may be desirable to have the server carry out specified tasks. For example, the server could generate individual similarity tables for the client or generate playlists for the client. This may be of greater use for the client to carry out these tasks for the portable media player which typically lacks the processing resources desired for large processes. In another example, it might be beneficial for the server to carry out some of the constraints used in playlist generation and incorporate the results in the similarity table data. For example, in the case of a genre constraint which will never be relaxed, the server may exclude media items from a non-compatible genre from an item's similarity data. In this way, the playlist module will not need to run the constraint when generating a playlist because the server would have already excluded items that would not pass the constraint from the individual similarity table.
  • [0164]
    FIG. 22 illustrates an example system for syncing media and similarity tables to a digital media player. The communication medium 920 is most likely to be the Internet, but can include any channel(s) allowing devices to communicate with each other. The online store 922 communicates with a client 924, typically a personal computer, with special software installed to communicate and interact with the online store. A portable media player 926, such as an Apple IPOD media player, typically syncs via a 30-pin USB connector with a personal computer 924 connected to the online store 922. Other ways to connect a portable media player 926 to a personal computer 924 include USB, Firewire, 802.11 series or Bluetooth wireless connections. The online store has a master library of media 930, a master similarity table 932, and a tree of server clusters 931. When the portable media player 926 is synced, the local media 934 metadata and the local subset similarity table 936 and participant server clusters 935 are transmitted through the personal computer 924 and the communication medium 920 to the online store 922. It is important to note that only metadata is transmitted, not the actual media. The online store does not need a copy of the media, it only needs sufficient information to positively identify which media are located on the portable media player 926. The online store is able to generate and/or retrieve the appropriate information and send it back to the device, updating the local subset similarity table 936 and participant clusters 935 with fresh data corresponding to the locally stored media 934. The local subset similarity table is then used to generate individualized playlists at the online store level, the computer level, and/or on the device itself.
  • [0165]
    In another embodiment the portable media player 926 can obtain data directly from the server 922. In such an embodiment the portable media player can be equipped with various capabilities for connecting to the online store such as through 802.11 series communications (Wi-Fi) or through a mobile telephone network. The portable media player can connect directly to a server and perform just as a client device would in the example above. The portable media can be assigned a unique identification number, or if the library is also associated with a client device, the portable media player can use the same identification number as would the client device. Regardless, the portable media player can upload information about media items stored in its library and download similarity data about those same items. It should be appreciated that the portable media player can have more than one method of connecting to the server. Not only can the portable media player have multiple mechanisms for connecting to an online store, but the portable media player can also have capabilities to synchronize with a client device and communicate directly with the online store.
  • [0166]
    Embodiments within the scope of the present invention may also include computer-readable media for carrying or having computer-executable instructions or data structures stored thereon. Such computer-readable media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code means in the form of computer-executable instructions or data structures. Combinations of the above should also be included within the scope of the computer-readable media.
  • [0167]
    Computer-executable instructions include, for example, instructions and data which cause a general purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Computer-executable instructions also include program modules that are executed by computers in stand-alone or network environments. Generally, program modules include routines, programs, objects, components, and data structures that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of the program code means for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represent examples of corresponding acts for implementing the functions described in such steps.
  • [0168]
    Those of skill in the art will appreciate that other embodiments of the invention may be practiced in network computing environments with many types of computer system configurations, including personal computers, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, and the like. Embodiments may also be practiced in distributed computing environments where tasks are performed by local and remote processing devices that are linked (either by hardwired links, wireless links, or by a combination thereof) through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.
  • [0169]
    Communication at various stages of the described system can be performed through a network cloud 328 (shown in FIG. 16) such as a local area network, a token ring network, the Internet, a corporate intranet, 802.11 series wireless signals, fiber-optic network, radio or microwave transmission, etc. Although the underlying communication technology may change, the fundamental principles described herein are still applicable.
  • [0170]
    The various embodiments described above are provided by way of illustration only and should not be construed to limit the invention. For example, the principles herein may be applied to an online store accessible wirelessly by a portable media playback device or by a personal computer physically connected to a network. Those skilled in the art will readily recognize various modifications and changes that may be made to the present invention without following the example embodiments and applications illustrated and described herein, and without departing from the true spirit and scope of the present disclosure.
Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US4996642 *25 Sep 198926 Feb 1991Neonics, Inc.System and method for recommending items
US5483278 *28 Sep 19939 Jan 1996Philips Electronics North America CorporationSystem and method for finding a movie of interest in a large movie database
US5616876 *19 Apr 19951 Apr 1997Microsoft CorporationSystem and methods for selecting music on the basis of subjective content
US5724521 *3 Nov 19943 Mar 1998Intel CorporationMethod and apparatus for providing electronic advertisements to end users in a consumer best-fit pricing manner
US5890152 *9 Sep 199630 Mar 1999Seymour Alvin RapaportPersonal feedback browser for obtaining media files
US6041311 *28 Jan 199721 Mar 2000Microsoft CorporationMethod and apparatus for item recommendation using automated collaborative filtering
US6047311 *14 Jul 19974 Apr 2000Matsushita Electric Industrial Co., Ltd.Agent communication system with dynamic change of declaratory script destination and behavior
US6345288 *15 May 20005 Feb 2002Onename CorporationComputer-based communication system and method using metadata defining a control-structure
US6346951 *23 Sep 199712 Feb 2002Touchtunes Music CorporationProcess for selecting a recording on a digital audiovisual reproduction system, for implementing the process
US6347313 *1 Mar 199912 Feb 2002Hewlett-Packard CompanyInformation embedding based on user relevance feedback for object retrieval
US6349339 *19 Nov 199919 Feb 2002Clickradio, Inc.System and method for utilizing data packets
US6381575 *11 Feb 200030 Apr 2002Arachnid, Inc.Computer jukebox and computer jukebox management system
US6526411 *15 Nov 200025 Feb 2003Sean WardSystem and method for creating dynamic playlists
US6532469 *20 Sep 199911 Mar 2003Clearforest Corp.Determining trends using text mining
US6687696 *26 Jul 20013 Feb 2004Recommind Inc.System and method for personalized search, information filtering, and for generating recommendations utilizing statistical latent class models
US6690918 *5 Jan 200110 Feb 2004Soundstarts, Inc.Networking by matching profile information over a data packet-network and a local area network
US6704576 *27 Sep 20009 Mar 2004At&T Corp.Method and system for communicating multimedia content in a unicast, multicast, simulcast or broadcast environment
US6727914 *17 Dec 199927 Apr 2004Koninklijke Philips Electronics N.V.Method and apparatus for recommending television programming using decision trees
US6842761 *8 Jun 200111 Jan 2005America Online, Inc.Full-text relevancy ranking
US6850252 *5 Oct 20001 Feb 2005Steven M. HoffbergIntelligent electronic appliance system and method
US6865546 *19 Apr 20008 Mar 2005Amazon.Com, Inc.Methods and systems of assisting users in purchasing items
US6987221 *30 May 200217 Jan 2006Microsoft CorporationAuto playlist generation with multiple seed songs
US6990497 *26 Jun 200124 Jan 2006Microsoft CorporationDynamic streaming media management
US6993532 *30 May 200131 Jan 2006Microsoft CorporationAuto playlist generator
US7013238 *15 Jul 200314 Mar 2006Microsoft CorporationSystem for delivering recommendations
US7020637 *30 Dec 200228 Mar 2006Timothy Robert BrattonDigital audio and video playback with performance complement testing
US7159000 *21 Mar 20022 Jan 2007Microsoft CorporationMethods and systems for repairing playlists
US7174126 *5 May 20036 Feb 2007Time Warner Interactive Video Group Inc.Technique for effectively accessing programming listing information in an entertainment delivery system
US7180473 *22 Feb 200220 Feb 2007Yokowo Co., Ltd.Antenna with built-in filter
US7194421 *12 Sep 200220 Mar 2007Erinmedia, LlcContent attribute impact invalidation method
US7196258 *21 Oct 200527 Mar 2007Microsoft CorporationAuto playlist generation with multiple seed songs
US7197472 *12 Sep 200227 Mar 2007Erinmedia, LlcMarket data acquisition system
US7340455 *27 Jan 20054 Mar 2008Microsoft CorporationClient-based generation of music playlists from a server-provided subset of music similarity vectors
US7345232 *6 Nov 200318 Mar 2008Nokia CorporationAutomatic personal playlist generation with implicit user feedback
US7478323 *11 Mar 200513 Jan 2009Apple Inc.Automated creation of media asset illustration collage
US7487107 *26 Jun 20073 Feb 2009International Business Machines CorporationMethod, system, and computer program for determining ranges of potential purchasing amounts, indexed according to latest cycle and recency frequency, by combining re-purchasing ratios and purchasing amounts
US7490775 *30 Dec 200417 Feb 2009Aol Llc, A Deleware Limited Liability CompanyIntelligent identification of multimedia content for synchronization
US7493572 *22 Jul 200417 Feb 2009Xerox CorporationNavigation methods, systems, and computer program products for virtual three-dimensional books
US7499630 *30 Mar 20023 Mar 2009Robert Bosch GmbhMethod for playing back multimedia data using an entertainment device
US7505959 *19 Nov 200417 Mar 2009Microsoft CorporationSystem and methods for the automatic transmission of new, high affinity media
US7644077 *21 Oct 20045 Jan 2010Microsoft CorporationMethods, computer readable mediums and systems for linking related data from at least two data sources based upon a scoring algorithm
US7647613 *21 Jul 200512 Jan 2010Akoo International, Inc.Apparatus and method for interactive content requests in a networked computer jukebox
US7650570 *4 Oct 200619 Jan 2010Strands, Inc.Methods and apparatus for visualizing a music library
US7653761 *15 Mar 200626 Jan 2010Microsoft CorporationAutomatic delivery of personalized content to a portable media player with feedback
US7657224 *6 May 20032 Feb 2010Syncronation, Inc.Localized audio networks and associated digital accessories
US7657493 *20 Dec 20062 Feb 2010Microsoft CorporationRecommendation system that identifies a valuable user action by mining data supplied by a plurality of users to find a correlation that suggests one or more actions for notification
US7680849 *25 Oct 200416 Mar 2010Apple Inc.Multiple media type synchronization between host computer and media device
US7680959 *11 Jul 200616 Mar 2010Napo Enterprises, LlcP2P network for providing real time media recommendations
US7875788 *10 May 200725 Jan 2011Harman International Industries, IncorporatedHeuristic organization and playback system
US7889724 *13 Apr 200715 Feb 2011Wideorbit, Inc.Multi-station media controller
US8260656 *19 Apr 20024 Sep 2012Amazon.Com, Inc.Mining of user-generated playlists for data regarding relationships between digital works
US20020002899 *22 Mar 200010 Jan 2002Gjerdingen Robert O.System for content based music searching
US20020042912 *3 Jan 200111 Apr 2002Jun IijimaPersonal taste profile information gathering apparatus
US20030022953 *11 Jul 200230 Jan 2003Shipley Company, L.L.C.Antireflective porogens
US20030033321 *23 Oct 200113 Feb 2003Audible Magic, Inc.Method and apparatus for identifying new media content
US20030055689 *2 Aug 200220 Mar 2003David BlockAutomated internet based interactive travel planning and management system
US20040002993 *26 Jun 20021 Jan 2004Microsoft CorporationUser feedback processing of metadata associated with digital media files
US20040003392 *26 Jun 20021 Jan 2004Koninklijke Philips Electronics N.V.Method and apparatus for finding and updating user group preferences in an entertainment system
US20040068552 *26 Dec 20018 Apr 2004David KotzMethods and apparatus for personalized content presentation
US20040070538 *22 Feb 200215 Apr 2004Ryo HorieAntenna incorporating filter
US20040073924 *30 Sep 200215 Apr 2004Ramesh PendakurBroadcast scheduling and content selection based upon aggregated user profile information
US20050004941 *24 Oct 20026 Jan 2005Maria Kalker Antonius Adrianus CornelisFingerprint database updating method, client and server
US20050019114 *25 Jul 200327 Jan 2005Chien-Min SungNanodiamond PCD and methods of forming
US20050021470 *8 Jun 200427 Jan 2005Bose CorporationIntelligent music track selection
US20050050079 *5 Oct 20043 Mar 2005Microsoft CorporationMethods and systems for per persona processing media content-associated metadata
US20050060350 *15 Sep 200317 Mar 2005Baum Zachariah JourneySystem and method for recommendation of media segments
US20050075908 *23 Nov 20047 Apr 2005Dian StevensPersonal business service system and method
US20060015571 *1 Jul 200519 Jan 2006International Business Machines CorporationComputer evaluation of contents of interest
US20060015904 *16 Jun 200519 Jan 2006Dwight MarcusMethod and apparatus for creation, distribution, assembly and verification of media
US20060018208 *21 Jul 200526 Jan 2006Guy NathanDigital downloading jukebox system with central and local music servers
US20060018209 *21 Jul 200526 Jan 2006Niko DrakoulisApparatus and method for interactive content requests in a networked computer jukebox
US20060020062 *8 Jul 200526 Jan 2006Bloom Paul DEpoxidized esters of vegetable oil fatty acids as reactive diluents
US20060020662 *19 Sep 200526 Jan 2006Emergent Music LlcEnabling recommendations and community by massively-distributed nearest-neighbor searching
US20060026263 *27 Sep 20052 Feb 2006Microsoft CorporationManaging stored data on a computer network
US20060032363 *21 Oct 200516 Feb 2006Microsoft CorporationAuto playlist generation with multiple seed songs
US20060053077 *14 Oct 20059 Mar 2006International Business Machines CorporationDigital content distribution using web broadcasting services
US20060059225 *31 Mar 200516 Mar 2006A9.Com, Inc.Methods and apparatus for automatic generation of recommended links
US20060062094 *9 Sep 200523 Mar 2006Guy NathanDigital downloading jukebox system with user-tailored music management, communications, and other tools
US20060067296 *1 Aug 200530 Mar 2006University Of WashingtonPredictive tuning of unscheduled streaming digital content
US20060195516 *28 Feb 200631 Aug 2006Yahoo! Inc.Method and system for generating affinity based playlists
US20070016507 *12 Jul 200518 Jan 2007Wally TzaraSystem for analysis and prediction of financial and statistical data
US20070043829 *17 Aug 200522 Feb 2007Robin DuaMethod and system for accessing a storage or computing device via the Internet
US20070073529 *29 Sep 200529 Mar 2007International Business Machines CorporationApparatus and method for composite behavioral modeling for multiple-sourced integrated circuits
US20080004990 *28 Jun 20063 Jan 2008Microsoft CorporationVirtual spot market for advertisements
US20080010266 *10 Jul 200610 Jan 2008Brunn Jonathan FA Context-Centric Method of Automated Introduction and Community Building
US20080021851 *6 Jul 200724 Jan 2008Music Intelligence SolutionsMusic intelligence universe server
US20080027881 *7 Jul 200631 Jan 2008Stephan Kurt Jan BisseMarket trader training tool
US20080040326 *14 Aug 200614 Feb 2008International Business Machines CorporationMethod and apparatus for organizing data sources
US20080046317 *21 Aug 200621 Feb 2008The Procter & Gamble CompanySystems and methods for predicting the efficacy of a marketing message
US20080065659 *10 Sep 200713 Mar 2008Akihiro WatanabeInformation processing apparatus, method and program thereof
US20080077264 *20 Sep 200727 Mar 2008Google Inc.Digital Audio File Management
US20090006353 *3 May 20051 Jan 2009Koninklijke Philips Electronics, N.V.Method and Apparatus for Selecting Items from a Number of Items
US20090024504 *2 May 200822 Jan 2009Kevin LermanSystem and method for forecasting fluctuations in future data and particularly for forecasting security prices by news analysis
US20090024510 *18 Jul 200722 Jan 2009Yahoo! Inc.System and method of making markets for a finite subset of orders placed across continuous and countably infinite outcome spaces
US20090043811 *12 Jun 200612 Feb 2009Noriyuki YamamotoInformation processing apparatus, method and program
US20090048957 *2 Apr 200819 Feb 2009Matthew CelanoMethod and system for financial counseling
US20090070267 *12 May 200612 Mar 2009Musicstrands, Inc.User programmed media delivery service
US20090073174 *13 Sep 200719 Mar 2009Microsoft CorporationUser interface for expressing forecasting estimates
US20090076939 *13 Sep 200719 Mar 2009Microsoft CorporationContinuous betting interface to prediction market
US20090076974 *13 Sep 200719 Mar 2009Microsoft CorporationCombined estimate contest and prediction market
US20090083307 *22 Apr 200526 Mar 2009Musicstrands, S.A.U.System and method for acquiring and adding data on the playing of elements or multimedia files
Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US8190663 *6 Jul 200929 May 2012Osterreichisches Forschungsinstitut Fur Artificial Intelligence Der Osterreichischen Studiengesellschaft Fur Kybernetik Of FreyungMethod and a system for identifying similar audio tracks
US8402027 *11 Feb 201019 Mar 2013Disney Enterprises, Inc.System and method for hybrid hierarchical segmentation
US8886649 *19 Mar 201211 Nov 2014Microsoft CorporationMulti-center canopy clustering
US8949237 *6 Jan 20123 Feb 2015Microsoft CorporationDetecting overlapping clusters
US8990208 *22 Sep 201124 Mar 2015Fujitsu LimitedInformation management and networking
US9009156 *10 Nov 200914 Apr 2015Hrl Laboratories, LlcSystem for automatic data clustering utilizing bio-inspired computing models
US9189802 *23 Aug 201217 Nov 2015Seoul National University R&Db FoundationMethod for contents recommendation
US931718524 Apr 201419 Apr 2016Apple Inc.Dynamic interactive entertainment venue
US9324158 *16 Sep 201326 Apr 2016Casio Computer Co., Ltd.Image processing device for performing image processing on moving image
US937307823 Apr 201221 Jun 2016Anametrix, Inc.Methods and systems for predictive alerting
US939588315 Mar 201319 Jul 2016Anametrix, Inc.Systems and method for integration of business analytics and business networking
US9489352 *13 May 20128 Nov 2016Groupon, Inc.System and method for providing content to users based on interactions by similar other users
US9734235 *18 Oct 201415 Aug 2017Microsoft Technology Licensing, LlcGrouping documents and data objects via multi-center canopy clustering
US20110004642 *6 Jul 20096 Jan 2011Dominik SchnitzerMethod and a system for identifying similar audio tracks
US20120272168 *20 Apr 201125 Oct 2012PanafoldMethods, apparatus, and systems for visually representing a relative relevance of content elements to an attractor
US20130080428 *22 Sep 201128 Mar 2013Fujitsu LimitedUser-Centric Opinion Analysis for Customer Relationship Management
US20130246429 *19 Mar 201219 Sep 2013Microsoft CorporationMulti-center canopy clustering
US20130325609 *4 Jun 20135 Dec 2013Microsoft CorporationManagement of media content availability
US20130339154 *15 Jun 201219 Dec 2013Apple Inc.System and method for assembling personalized offers
US20140059128 *23 Aug 201227 Feb 2014Kibeom LeeMethod for Contents Recommendation
US20140079290 *16 Sep 201320 Mar 2014Casio Computer Co., Ltd.Image processing device for performing image processing on moving image
US20140114772 *7 Jun 201324 Apr 2014Apple Inc.Personalized media stations
US20140129571 *15 Mar 20138 May 2014Axwave Inc.Electronic media signature based applications
US20150039619 *18 Oct 20145 Feb 2015Microsoft CorporationGrouping documents and data objects via multi-center canopy clustering
US20150046458 *16 Mar 201312 Feb 2015Tencent Technology (Shenzhen) Company LimitedMethod for recommending users in social network and the system thereof
US20150074572 *14 Nov 201412 Mar 2015Frederic SigalNavigable wall
US20150135048 *23 Jan 201514 May 2015PanafoldMethods, apparatus, and systems for visually representing a relative relevance of content elements to an attractor
US20150256885 *26 May 201510 Sep 2015Thomson LicensingMethod for determining content for a personal channel
US20170068731 *17 Nov 20169 Mar 2017Axwave Inc.Electronic media signature based applications
EP2895971A4 *19 Aug 20134 May 2016Gracenote IncUser profile based on clustering tiered descriptors
Classifications
U.S. Classification707/737, 707/E17.101, 707/E17.046, 707/E17.032
International ClassificationG06F17/30
Cooperative ClassificationG06F17/30749, G06F17/30772, G06F17/30766
European ClassificationG06F17/30U4P, G06F17/30U3F2, G06F17/30U2
Legal Events
DateCodeEventDescription
23 Dec 2009ASAssignment
Owner name: APPLE INC., CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GATES, PATRICK;WERNER, JEREMY;VYRROS, ANDREW H.;AND OTHERS;SIGNING DATES FROM 20091125 TO 20091217;REEL/FRAME:023701/0240