JP2006048287A - Information processing device and method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To widely recommend contents such as contents that a user is inexperienced in. <P>SOLUTION: A matrix generation part 18 generates a meta-data matrix consisting of rows corresponding to N meta-data and columns corresponding to M contents. An LSA arithmetic part 20 subjects the meta-data matrix to singular value decomposition processing to generate its approximate matrix. A vector arithmetic part 22 clusters M column components of the approximate matrix into S clusters to generate an UPV for every S clusters. The vector arithmetic part 22 selects one or more groups of representative vectors of two clusters out of representative vectors of S clusters and generates a difference UPV between UPVs of the two clusters as to the selected one or more groups. A content recommendation part 23 recommends contents by using the difference UPV. This invention is applicable to an information processor recommending contents. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an information processing apparatus and method, and a program, and in particular, realizes a wider range of content recommendation and content recommendation that the user has never experienced compared to conventional content recommendation simply using UPV. The present invention relates to an information processing apparatus and method, and a program.

  In recent years, as one of information processing systems, a system for recommending content to users (hereinafter referred to as a content recommendation system) has become widespread.

  Hereinafter, an outline of a series of processing (hereinafter referred to as content recommendation processing) until the conventional content recommendation system recommends content will be described.

  However, for simplicity of explanation, it is assumed that one information processing apparatus executes all of the content recommendation processing.

  First, the information processing apparatus vectorizes the content using the metadata assigned to the content as a base vector. Hereinafter, such a vector is referred to as a content vector.

  Next, the information processing apparatus generates a plurality of such content vectors and arranges each of the plurality of content vectors in order in a predetermined direction, that is, a matrix having a plurality of content vectors as row components or column components. Is generated. Hereinafter, such a matrix is referred to as a metadata matrix. A space spanned by all metadata using the metadata as a base vector is referred to as a metadata space.

  Next, the information processing apparatus weights each component of the metadata matrix using a predetermined weighting method (gives a weight value). As the weighting method, for example, the appearance frequency in the metadata content, the TF / IDF method based on the completeness and specificity in the metadata content (frequency of metadata included in the content: TF, and its A weighting method using a reciprocal of the number of contents including metadata: a method using an IDF multiplied) is widely used.

  In this manner, each column component or each row component of the metadata matrix, that is, the content vector becomes a content vector appropriately weighted by the metadata.

  Furthermore, the information processing apparatus generates a vector indicating the user's preference by using one or more weighted content vectors as necessary. Hereinafter, a vector indicating the user's preference is referred to as a user preference vector or UPV (User Preference Vector).

  Then, the information processing apparatus obtains similarities such as cosine correlation between the UPV and feature vectors corresponding to each of the plurality of contents that the user has not experienced (performs matching processing), for example, in descending order of similarity. Recommend content.

  The outline of the content recommendation process in the conventional content recommendation system has been described above.

  By the way, in recent years, a technique for performing matching in a dimension-compressed space using a technique called LSA (Latent Semantci Analysis) has been established (see Non-Patent Documents 1 to 3 and Patent Document 1). Such a technique using LSA has a track record as a technique related to document classification and retrieval in consideration of semantic grouping between words.

  It is also possible to apply such a technique using LSA to content recommendation processing.

  In other words, when the information processing apparatus performs singular value decomposition on the above-described metadata matrix, as a result, a conceptual space in which highly related metadata are grouped into one dimension from the above-described metadata space. Generated. Each base in the concept space is associated with a singular value (the importance of the base). Therefore, when the information processing device uses only the upper base with a large singular value (dimension compression) and performs the back projection to the metadata space, as a result, a matrix in which the relationship between the metadata is revealed Generated. Hereinafter, such a matrix is referred to as an approximate matrix.

The above-described series of processing is called LSA, and the information processing apparatus can also perform content matching processing by using the approximate matrix generated by LSA instead of the metadata matrix.
U.S. Pat. US Patent No. 5301109 SC Deerwester, ST Dumais, TK Landauer, GW Furnas, and RA Harshman. “Indexing by latent semantic analysis.” Journal of the American Society of Information Science, 41 (6): 391-407, 1990. Japanese Patent Laid-Open No. 11-296552

  However, as described above, the UPV (user preference vector) generation method of the conventional content recommendation system based on the vector space method is used to generate the UPV based on the average of the content vectors of the content group that the user gave high evaluation. In many cases, a generation method is employed. The UPV generated by such a generation method is a vector that imitates the user's various preferences, and even if content recommendation is performed using such UPV, it is difficult to make a wide range of recommendations was there. In addition, even if a highly evaluated content group is clustered into a plurality of groups to produce a variety, there is a problem that it is difficult to recommend content that the user has never experienced.

  The present invention has been made in view of such a situation. Compared to the conventional content recommendation that simply uses UPV, the present invention recommends a wider range of content recommendation and a content recommendation not experienced by the user. It can be realized.

  The information processing apparatus of the present invention is based on N (N is an integer value of 1 or more) metadata associated with at least one of a plurality of contents, and M (M is A matrix generating unit that vectorizes each of the contents of (an integer value of 1 or more) and generates a matrix having the resulting M vectors as column components or row components as a metadata matrix, and generated by the matrix generating unit By performing singular value decomposition on the metadata matrix, an approximate matrix generating means for generating an approximate matrix of the metadata matrix and each of M contents of the approximate matrix generated by the approximate matrix generating means are shown. Each of the M column components or row components is partitioned as a vector, and each of the M partitioned vectors is divided into S (S is an integer value less than or equal to M) clusters. A representative vector representing the corresponding cluster based on the vector classified into the corresponding cluster by the classification means for each of the S classification clusters and the S clusters, A representative vector generating means for generating a representative vector based on the metadata of the first and a representative vector generating means for selecting one or more pairs of representative vectors of two clusters from among the representative vectors of the S clusters generated by the representative vector generating means. For each of one or more selected sets, difference vector generation means for generating a difference vector of representative vectors of two clusters, and other contents different from M contents are vectorized based on N metadata Of the resulting vector and one or more difference vectors generated by the difference vector generation means Characterized in that it comprises a similarity calculation means for calculating at least one of similarity.

  Based on at least one of the one or more similarities calculated by the similarity calculation means, the other content is determined by the recommendation means and the recommendation means determines whether or not the content should be recommended to the user. When it is determined that the content should be recommended to the user, it is possible to further provide a presentation means for presenting the other content to the user.

  The information processing method of the present invention is based on N (N is an integer value of 1 or more) metadata associated with at least one of a plurality of contents, and M (M is A matrix generation step for vectorizing each of the contents having an integer value of 1 or more) and generating a matrix having the resulting M vectors as column components or row components as a metadata matrix, and generation by processing of the matrix generation step An approximate matrix generation step for generating an approximate matrix of the metadata matrix by performing singular value decomposition on the generated metadata matrix, and M contents of the approximate matrix generated by the processing of the approximate matrix generation step Each of the column component or the row component indicating each of the vectors is partitioned as a vector, and each of the M partitioned vectors is divided into S (S is equal to or less than M A classifying step for classifying a predetermined one of the clusters of (integer value of) and representing the corresponding cluster for each of the S clusters based on the vectors classified into the corresponding cluster by the processing of the classifying step. A representative vector generation step of generating a representative vector based on N pieces of metadata, and two representative vectors of S clusters generated by the processing of the representative vector generation step. A difference vector generation step of selecting one or more sets of cluster representative vectors, and generating a difference vector of representative vectors of two clusters for each of the selected one or more sets, and other contents different from the M contents Is vectorized using N metadata as a basis, and the resulting vector and the difference vector Characterized in that it comprises a similarity calculation step of calculating at least one of similarity of the one or more difference vectors generated by the processing of Le generating step.

  The program of the present invention is a program to be executed by a computer, and is based on N (N is an integer value of 1 or more) metadata associated with at least one of the plurality of contents. A matrix generation step of vectorizing each of M contents (M is an integer value of 1 or more), and generating a matrix having the resulting M vectors as column components or row components as a metadata matrix; An approximate matrix generation step for generating an approximate matrix of the metadata matrix by performing singular value decomposition on the metadata matrix generated by the matrix generation step processing, and an approximation matrix generated by the processing of the approximate matrix generation step Each of the column component or row component indicating each of the M contents is divided into a vector and divided into M pieces. A classification step for classifying each of the vectors into a predetermined one of S clusters (S is an integer value less than or equal to M), and a cluster corresponding to each of the S clusters by processing of the classification step Based on the vectors classified into the above, the representative vector representing the corresponding cluster, which is generated by the representative vector generation step for generating the representative vector based on the N pieces of metadata and the processing of the representative vector generation step. A difference vector that selects one or more pairs of representative vectors of two clusters from the representative vectors of S clusters and generates a difference vector of representative vectors of the two clusters for each of the selected one or more sets. Generation step and other content different from M content vectors based on N metadata And, to the resulting and vector, characterized in that it comprises a similarity calculation step of calculating at least one of similarity of the one or more difference vectors generated by the processing of the difference vector generation step.

  In the information processing apparatus, method, and program of the present invention, a plurality of contents are based on N (N is an integer value of 1 or more) metadata associated with at least one of the plurality of contents. Each of the M contents (M is an integer value of 1 or more) is vectorized, and a matrix having the resulting M vectors as column components or row components is generated as a metadata matrix. In addition, singular value decomposition is performed on the metadata matrix to generate an approximate matrix of the metadata matrix. Next, each of the column component or row component indicating each of the M contents in the approximate matrix is partitioned as a vector, and each of the M partitioned vectors is S (S is an integer equal to or less than M). A representative vector representing a corresponding cluster based on a vector classified into a corresponding cluster for each of the S clusters, and N A representative vector based on each piece of metadata is generated. Then, one or more pairs of representative vectors of two clusters are selected from the representative vectors of S clusters, and a difference vector of the representative vectors of the two clusters is generated for each of the selected one or more sets. . Then, other contents different from the M contents are vectorized based on the N metadata, and the similarity between the resulting vector and at least one of the one or more difference vectors is calculated. .

  As described above, according to the present invention, content metadata can be handled. In particular, when the content vectors are clustered and the representative vector of each cluster is generated as, for example, a UPV, a difference vector of UPVs of two clusters (a difference UPV described later) can be generated. Therefore, by calculating the degree of similarity using this difference vector, a wider range of content recommendations and a user inexperienced content recommendation are compared with the conventional content recommendation that simply uses UPV. realizable.

  Embodiments of the present invention will be described below. Correspondences between constituent elements described in the claims and specific examples in the embodiments of the present invention are exemplified as follows. This description is to confirm that specific examples supporting the invention described in the claims are described in the embodiments of the invention. Therefore, even though there are specific examples that are described in the embodiment of the invention but are not described here as corresponding to the configuration requirements, the specific examples are not included in the configuration. It does not mean that it does not correspond to a requirement. On the contrary, even if a specific example is described here as corresponding to a configuration requirement, this means that the specific example does not correspond to a configuration requirement other than the configuration requirement. not.

  Further, this description does not mean that all the inventions corresponding to the specific examples described in the embodiments of the invention are described in the claims. In other words, this description is an invention corresponding to the specific example described in the embodiment of the invention, and the existence of an invention not described in the claims of this application, that is, in the future, a divisional application will be made. Nor does it deny the existence of an invention added by amendment.

  According to the present invention, an information processing apparatus is provided. This information processing apparatus (the information processing apparatus of FIG. 1 and the information processing apparatus of the third embodiment described later) is associated with at least one of a plurality of contents (N is an integer of 1 or more). Each of the M contents (M is an integer value of 1 or more) of the plurality of contents is vectorized based on the (numerical) metadata, and the resulting M vectors are column components or row components. Matrix generating means for generating a matrix to be processed as a metadata matrix (for example, the matrix generating section 18 in FIG. 1 (FIG. 9), the weighting processing section 19 may be added), and the metadata generated by the matrix generating means By performing singular value decomposition on the matrix, approximate matrix generation means (for example, the LSA operation unit 20 in FIG. 1 (FIG. 9)) that generates an approximate matrix of the metadata matrix, and the approximate matrix generation means Each of M column components or row components indicating each of the M pieces of the contents of the generated approximate matrix is partitioned as a vector, and each of the M partitioned vectors is defined as S (S Is a classifying means for classifying into a predetermined one of clusters having an integer value less than or equal to M (the vector calculation unit (clustering unit) 22 in FIG. 1 (FIG. 9) executing the processing of step S44 in FIG. 10) and , For each of the S clusters, a representative vector representing the corresponding cluster based on the vector classified into the corresponding cluster by the classification means, the N based on the metadata Representative vector generation means for generating a representative vector (vector operation unit (clustering unit) 22 in FIG. 1 (FIG. 9) that executes the processing of step S45 in FIG. 10) , One or more representative vector sets of two clusters are selected from the representative vectors of the S number of clusters generated by the representative vector generating means, and two clusters are selected for each of the selected one or more sets. Difference vector generation means for generating a difference vector (for example, a difference UPV described later) of the representative vector (vector operation unit (clustering unit) 22 of FIG. 1 (FIG. 9) that executes the process of step S45 of FIG. 10); Another content different from the M content is vectorized based on the N metadata, and the resulting vector and at least one of the one or more difference vectors generated by the difference vector generation means A difference vector generator for generating a similarity calculation means (for example, a differential UPV described later) for calculating the similarity with one Characterized in that it comprises (vector arithmetic unit of FIG. 1 which executes a process at step S47 in FIG. 10 (9) (matching section) 22) and.

  The information processing apparatus determines whether or not the other content is content to be recommended to a user based on at least one of the one or more similarities calculated by the similarity calculating means. Recommending means (for example, the content recommending unit 23 in FIG. 1 (FIG. 9)) and the recommending means determine that the other content should be recommended to the user, the other content is selected by the user. Presenting means (for example, the user interface unit 11 in FIG. 1 (FIG. 9)) to be presented can be further provided.

  According to the present invention, an information processing method is provided. A method corresponding to this information processing method (for example, the “recommendation process using a difference between clustered UPV groups” in FIG. 10) is N (N is 1 or more) associated with at least one of a plurality of contents. Of the plurality of contents (M is an integer value of 1 or more), and the resulting M vectors are column components or rows. A matrix generation step for generating a matrix as a component as a metadata matrix (for example, the process of step S41 in FIG. 10 may include the process of step S42), and the metadata generated by the process of the matrix generation step An approximate matrix generating step for generating an approximate matrix of the metadata matrix by performing singular value decomposition on the matrix (for example, step S4 in FIG. 10). 3) and the column component or the row component indicating each of the M contents of the approximate matrix generated by the process of the approximate matrix generation step as a vector and divided into M elements. A classification step (for example, the process of step S44 in FIG. 10) for classifying each of the vectors into a predetermined one of S clusters (S is an integer value less than or equal to M), and S For each of the clusters, a representative vector representing the corresponding cluster based on the vector classified into the corresponding cluster by the processing of the classification step, wherein the representative vector based on N pieces of the metadata is A representative vector generation step to be generated (for example, the process of step S45 in FIG. 10) and the process of the representative vector generation step One or more sets of representative vectors of two clusters are selected from the representative vectors of the S number of clusters, and a difference vector of representative vectors of the two clusters is generated for each of the selected one or more sets. A difference vector generation step (for example, the process of step S46 of FIG. 10), and other contents different from the M pieces of content are vectorized based on the N pieces of metadata, and the resulting vector, A similarity calculation step (for example, the process of step S47 in FIG. 10) for calculating the similarity with at least one of the one or more difference vectors generated by the process of the difference vector generation step. To do.

  According to the present invention, a program is provided. This program is a program corresponding to the information processing method of the present invention described above, and is executed by, for example, the computer of FIG.

  As described above, in the present invention, content and its metadata are to be processed.

  What should be noted here is that the contents and metadata in the present invention, that is, the contents and metadata that can be processed in the present invention are broad concepts compared to the contents and metadata that are generally called. It is a point.

  In other words, the content in the present invention includes a television broadcast program, a movie, a photograph, a song, etc. (moving image, still image, audio, or a combination thereof) generally called a content, a document It is a broad concept that refers to all software or hardware that can be used by a user, such as goods (including articles) and conversation. However, when the content is an article (hardware), for example, data obtained by projecting the article onto a moving image or a still image is used as the content data.

  Here, when there is no need to distinguish content and content data individually, they are collectively referred to as content.

  The metadata in the present invention refers to the following information. That is, as described above, the content in the present invention includes not only general content but also a user's private document (for example, mail). Therefore, the metadata in the present invention does not simply indicate general metadata such as program metadata, but the content (a broad concept of content) in the present invention itself or a part thereof, or the attribute of the content is a word. It is a broad concept that refers to digitized information (numerical values are also regarded as one of worded information). In other words, if it is information indicating one or more characteristics of the content in the present invention, all of the information can be metadata.

  Specifically, for example, the content includes a web page, mail, a web bulletin board, a book, and the like in addition to the above-described television broadcast program, movie, and music.

  In this case, examples of the metadata type of the television broadcast program include broadcast time, performers, staff, genre, and channel. Examples of movie metadata include airing time, performers, staff, genres, distributors, and the like. Examples of music metadata include artist names, genres, musical instruments, rhythms, moods, and the like. Examples of web page metadata include creators, links, linked links, URLs (regions etc.), appearance words, and the like. Examples of mail metadata include sender / receiver, transmission / reception date and time, and appearance words. The types of metadata on the Web bulletin board include authors, publishers, publication dates, and appearance words in the case of books.

  Next, an exemplary embodiment of an information processing system to which the present invention is applicable, which can handle content and metadata in a broad sense as described above, will be described with reference to the drawings.

  FIG. 1 shows a functional configuration example of an information processing system to which the present invention is applied.

  As shown in FIG. 1, the information processing system includes a user interface unit 11 to an information transmission unit 24.

  The user interface unit 11 includes an output device for a user to experience content and an input device for the user to perform operations on the content. Specifically, examples of the output device include a display and a speaker. Examples of the input device include a keyboard, a mouse, a remote controller, and a touch panel.

  The user profile storage unit 12 stores information such as pointers (ID numbers and the like) to contents that the user has experienced in the past and evaluations on the pointers. This evaluation is input by the user using the user interface unit 11.

  Therefore, the other blocks refer to various types of information stored in the user profile storage unit 12 to read out desired content from the content storage unit 17 and read out metadata associated therewith from the metadata storage unit 18. be able to.

  The user dictionary storage unit 13 stores frequent metadata of content experienced by the user and unnecessary metadata. Details of frequent metadata and unnecessary metadata will be described later. In addition, the user dictionary storage unit 13 may store a weight for user-specific metadata. In the user dictionary storage unit 13, data is exchanged with the user interface unit 11, the content recommendation unit 16, the metadata extraction unit 22, and the like, and any addition, deletion, or reference of any number of arbitrary data is performed. Is freely possible.

  The general dictionary storage unit 14 stores, for example, all appearance metadata and unnecessary metadata common to all users as metadata common to users. Further, the general dictionary storage unit 14 may store weights for metadata common to users. The general dictionary storage unit 14 also exchanges data with the user interface unit 11, the content recommendation unit 16, the metadata extraction unit 22, and the like, and can freely add, delete, and refer to arbitrary data. To do.

  The content storage unit 15 stores content that can be provided to the user, for example, video, music, text, Web, and the like. The main function of the content storage unit 15 is a function of providing data to the content recommendation unit 18 in response to a request from the content recommendation unit 18. An identifier such as an ID number is added to each content stored in the content storage unit 15. In addition, it is assumed that the content storage unit 15 can freely add, delete, and refer to an arbitrary number of arbitrary contents.

  The metadata storage unit 16 stores metadata corresponding to the content stored in the content storage unit 15. Note that storing metadata not only stores the metadata itself but also associates an arbitrary number of metadata with each content that can be identified by an identifier such as the ID number described above. It also refers to storing each frequency of each metadata and each heuristically determined weight.

  Each of the user profile storage unit 12 to the metadata storage unit 16 described above is configured as an area of a memory such as a hard disk.

  On the other hand, each of the metadata acquisition unit 17 to the content recommendation unit 23 to be described below may be configured by software or hardware if possible. Or you may comprise by those combination.

  The metadata acquisition unit 17 acquires metadata to be stored in the metadata storage unit 16 described above and stores the metadata in the metadata storage unit 16. For example, when the content is a sentence, the metadata acquisition unit 15 extracts, for example, a word appearing in the sentence, analyzes the appearance frequency of the word, and each of the words and their appearance frequencies. Are stored in the metadata storage unit 16 in association with each other.

  The matrix generation unit 18 accumulates the above-described content vectors representing each of a plurality of contents, and generates a metadata matrix having each content vector as a column component, for example. Note that the matrix generation unit 18 does not perform processing such as weighting.

  The weighting processing unit 19 weights the metadata matrix generated by the matrix generating unit 18 using various algorithms such as TF / IDF. Note that the timing of the weighting process of the weighting processing unit 19 is not particularly limited, and may be before or after the LSA calculation process of the LSA calculation unit 20 described later.

  The LSA operation unit 20 performs an LSA operation on the metadata matrix generated by the matrix generation unit 18 or the metadata matrix in which each component is weighted by the weighting processing unit 19. Here, the LSA calculation refers to the following first to third processes.

  The first process is a process for generating a projection matrix by performing singular value decomposition.

  The second process is to project each column component of the metadata matrix, that is, each content vector (group), into the concept space by the projection matrix generated by the first process.

  The third process is a process for generating an approximate matrix of a metadata matrix, that is, an approximate matrix whose dimensions are appropriately compressed with respect to the metadata matrix.

  Hereinafter, the LSA calculation will be described in more detail.

  For example, suppose that a metadata matrix D with N rows and M columns is supplied from the matrix generation unit 18 or the weighting processing unit 19 to the LSA calculation unit 20.

  In this case, as one of the first processes, the LSA computing unit 20 performs singular value decomposition on the metadata matrix D of N rows and M columns, thereby satisfying the following equation (1) for the metadata matrix D. Decompose into component matrices U, Σ, and V. In Equation (1), the component matrix U is an N-row N-column left singular vector, the component matrix V is an M-row M-column right singular vector, and the component matrix Σ is an N-row M-column singular matrix. Show. V˜ represents a transposed matrix of the component matrix V.

  D = UΣV ~ (1)

  Here, if the rank of the metadata matrix D is r (r is an integer value less than or equal to N, M), the component matrix Σ has r singular values arranged as diagonal elements, and the other elements are all 0. Is a matrix of In addition, since the first r column components (left singular vectors) of the component matrix U are orthonormal basis and are important column components in order from the left, k (k is an integer value smaller than r) left The best approximation can be achieved by expressing (projecting) each content vector using a singular vector.

Therefore, as one of the first processes, the LSA arithmetic unit 20 performs a projection matrix (hereinafter referred to as U _k ) composed of the first k column components (left singular vectors) of the component matrix U, that is, N rows. A k-column projection matrix U _k is generated.

Next, as a second process, the LSA computing unit 20 applies the transposed matrix of the projection matrix U _k from the left to each column component of the metadata matrix D, that is, each content vector (N dimension). By multiplying, each content vector reduced in dimension to k dimensions (each approximate vector of each content vector) is generated. That is, the LSA computing unit 20 projects each content vector onto a k-dimensional conceptual space. In other words, it can be said that the LSA operation unit 20 generates the concept space by generating the projection matrix U _k in the first process.

In addition, as one of the third processes, the LSA arithmetic unit 20 similarly uses the first k right singular vectors for the component matrix V, and uses the first k column components (right singular) of the component matrix V. A matrix (hereinafter referred to as “V _k ”), that is, a matrix V _k having M rows and k columns is generated.

Further, as one of the third processes, the LSA arithmetic unit 20 performs the elements from 1 to k rows (k × k of the component matrix Σ) of the first k column components of the component matrix Σ. Matrix (hereinafter referred to as Σ _k ), that is, a matrix Σ _k having k rows and k columns.

Then, as one of the third processes, the LSA computing unit 20 computes the right side of the following equation (2) to generate an approximate matrix D _k whose rank is reduced to _k . In Equation (2), V _k ~ _represents a transposed matrix of the component matrix V _k .

D _k = U _k Σ _k V _k ~ (2)

  The LSA calculation executed by the LSA calculation unit 20 has been described above.

The metadata extraction unit 21 applies each component value of the metadata matrix D to which each component is weighted by the weighting processing unit 19 or each component value of the approximate matrix _Dk generated by the LSA operation of the LSA operation unit 20. Then, a predetermined calculation is performed, and characteristic metadata is extracted based on the calculation result. The metadata extraction unit 21 further notifies other blocks of the extracted metadata ID number and the like as appropriate.

The vector calculation unit 22 is a content vector group appropriately processed by the weighting processing unit 19 and the LSA calculation unit 20, that is, one or more column components of each column component of the metadata matrix D or the approximate matrix _Dk. A process (matching process) for calculating similarity between vectors by cosine correlation or the like, a clustering process for classifying into a plurality of groups, and the like are executed. Note that control of these processes is performed by the content recommendation unit 23.

The content recommendation unit 23 uses each component value of the metadata matrix D to which each component is weighted by the weighting processing unit 19 or the approximate matrix D _k generated by the LSA calculation of the LSA calculation unit 20 to use the vector calculation unit. Processing for requesting appropriate processing (matching processing and clustering processing described above) for 22, processing for reading predetermined content from the content storage unit 15, processing for presenting content to the user via the user interface unit 11, and the like are executed.

  The information transmission unit 24 transmits various information transmitted from a predetermined block of the user interface unit 11 to the content recommendation unit 23 described above to an appropriate block of the user interface unit 11 to the content recommendation unit 23. .

  The information processing system of the present invention has been described above with reference to FIG.

  When the information processing system of the present invention is composed of, for example, a client and a server, the user interface unit 11 of FIG. 1 is arranged in the client, but each of the other user profile storage unit 12 to content recommendation unit 23 is It may be arranged on the server side or on the client side.

  In this case, the information transmission unit 24 includes a communication device that communicates with another information processing apparatus via the network, and the communication device is provided in each of the server and the client. That is, the server and the client communicate with each other via a network using a built-in communication device.

  Further, in this case, the information transmission unit 24 may include various buses provided in each of the server and the client. That is, when at least two blocks of the user interface unit 11 to the content recommendation unit 23 are arranged in the client, information exchange between these blocks is performed via various buses in the client. Similarly, when at least two blocks of the user profile storage unit 12 to the content recommendation unit 23 are arranged in the server, information exchange between these blocks is performed via various buses in the server. Is called.

  Specifically, for example, the user interface unit 11, the user profile storage unit 12 related to user privacy, and the user dictionary storage unit 13 are arranged on the client side, and the other general dictionary storage unit 14 to content recommendation unit 23. Can be placed on the server side.

  Further, for example, the content storage unit 15 and the metadata storage 16 that require a large amount of storage capacity are arranged on the server side, and other blocks, that is, the user interface unit 11 to the general dictionary storage unit 14, and the metadata acquisition The unit 17 to the content recommendation unit 23 can be arranged on the client side.

  Further, for example, each of the user interface unit 11 to the content recommendation unit 23 can be appropriately distributed on the server side and the client side so as to distribute the calculation load.

  Further, for example, it is possible to take a form in which all of the user interface unit 11 to the content recommendation unit 23 are arranged on the client side. That is, all of the user interface unit 11 to the content recommendation unit 23 may be arranged in one information processing apparatus.

The information processing system of FIG. 1 having such a configuration, as described above, vectorizes each of various contents as corresponding metadata as a base vector, and a metadata matrix having each content vector obtained as a result, for example, as a column component D can be generated. Furthermore, the information processing system in FIG. 1 can perform weighting processing or LSA calculation on the metadata. This makes it possible to obtain a metadata matrix D in which each component is weighted and an approximate matrix _Dk thereof.

Therefore, the information processing system in FIG. 1 can perform various processes using the metadata matrix D weighted with each component and the approximate matrix _Dk . For example, the information processing system in FIG. 1 can execute processes invented by the present inventor, such as the following first to fifth processes.

  That is, it can be said that the inventor newly invented an information processing system or an information processing apparatus capable of executing each of the following first to fifth processes. And it can be said that this inventor disclosed the information processing system of the structure of FIG. 1 as one Embodiment of the structure. Accordingly, it is needless to say that the form is not limited to the example of FIG. 1 as long as it is an information processing system or information processing apparatus capable of executing each of the following first to fifth processes.

  The first process is “unnecessary metadata extraction process considering co-occurrence relation”. The second process is a “recommendation process considering the co-occurrence relationship”. The third process is a “recommendation process using a difference between clustered UPV (user preference vectors) groups”. The fourth processing is “content re-evaluation processing by LSA”. The fifth processing is “recommendation processing by hybrid of LSA and other methods”.

  Hereinafter, the details of the first to fifth processes will be individually described in that order. That is, hereinafter, an embodiment of an information processing system or an information processing apparatus that executes each of the first to fifth processes will be described individually in that order. In the following description, each of the embodiments of the information processing system or the information processing apparatus that executes each of the first to fifth processes will be described for the sake of simplicity. Called.

  (First embodiment)

  First, the first embodiment will be described.

  For example, when the content is a sentence, the frequency of words appearing in the document (or an appropriate weight value corresponding thereto) can be adopted as the metadata.

  In this case, when a new document to be processed is added, new words that have not appeared so far among the words that appear in the new sentence are meta-data as new metadata base vectors. Added to the data space.

  That is, the number of dimensions of the metadata space, that is, the number of dimensions of the content vector is equal to the number of types of words that have appeared in all sentences that have been processed so far. Therefore, as the number of sentences to be processed increases, that is, as the number of sentences created or browsed by the user increases, the number of dimensions in the metadata space also increases. Specifically, the number of dimensions of the metadata space is generally several thousand to several tens of thousands.

  As a result, there is a problem that subsequent calculations such as matching processing and clustering processing become difficult. At this time, attempts have been made to reduce the number of words based on word weights in the past, but when using TF / IDF, etc., the co-occurrence relationship (or synonym) of metadata (words) is considered However, there is a problem that words that should not be deleted are often deleted.

  In order to solve this problem, the present inventor has invented the above-described first processing, that is, “unnecessary metadata extraction processing considering co-occurrence relationships”.

This first process uses an approximate matrix D _k generated by LSA. This is because the approximate matrix D _k is a matrix generated in consideration of the co-occurrence relationship. However, the relationship between the approximate matrix D _k and the co-occurrence relationship will be described later.

  Hereinafter, with reference to FIG. 2 to FIG. 6, the information processing system or information processing apparatus of the first embodiment, that is, the information processing system or information processing apparatus that executes “unnecessary metadata extraction processing considering co-occurrence relationships” Will be described.

  FIG. 2 illustrates a functional configuration example of the information processing system or the information processing apparatus according to the first embodiment.

  In other words, blocks necessary for the execution of “unnecessary metadata extraction processing considering co-occurrence relation” are extracted from all the blocks of the user interface unit 11 to the content recommendation unit 23 of FIG. FIG. 2 is a diagram arranged according to the flow of information when executing the “unnecessary metadata extraction process considering co-occurrence relationships”. Therefore, the description of each block shown in FIG. 2 has been described above with reference to FIG.

  Although omitted in the example of FIG. 2, the information transmission unit 24 of FIG. 1 is actually arranged in each arrow connecting the two blocks, that is, between the two blocks. become.

  FIG. 3 is a flowchart for explaining an example of “unnecessary metadata extraction processing considering the co-occurrence relationship”. Therefore, an example of “unnecessary metadata extraction processing considering co-occurrence relations” will be described below with reference to the flowchart of FIG.

  In order to facilitate understanding of the “unnecessary metadata extraction process considering the co-occurrence relationship”, the following description will be given with reference to FIGS. 4 to 6 as appropriate. That is, FIGS. 4 to 6 show specific examples of processing results of “unnecessary metadata extraction processing considering co-occurrence relationships”.

  In step S <b> 1 of FIG. 3, the matrix generation unit 18 generates a metadata matrix D.

  Specifically, in step S <b> 1, the matrix calculation unit 18 acquires, from the user profile storage unit 12, each pointer (ID number or the like) to one or more contents that the user has experienced. Next, the matrix calculation unit 18 acquires metadata to which each pointer is attached, that is, metadata corresponding to each content that the user has experienced from the metadata storage unit 16, and acquires the acquired metadata. Using the data as a basis vector, each content that the user has experienced is vectorized. Thereby, a content vector corresponding to each content that the user has experienced is generated. Then, the matrix calculation unit 18 generates a metadata matrix D having each content vector as a column component.

  Note that the metadata accumulated in the process of step S1 may be metadata corresponding to all content in addition to metadata corresponding to content experienced by a single user, or corresponding to content experienced by a plurality of users. The metadata to be used may be used. However, the registration destination of unnecessary metadata in step S6 to be described later differs depending on the content to be processed in the metadata acquisition process.

  In step S2, the weighting processing unit 19 performs weighting on the metadata matrix D generated by the matrix generation unit 18 in the processing of step S1, using a predetermined weighting method.

  As described above, the weighting method used in the process of step S2 is not particularly limited, and the method using TF / IDF, the method using normalized TF, or the passage of time for each content or metadata is reflected. A heuristic weighting method or the like can be employed.

  However, here, for example, five sentences d1 to d5 are processed as contents, words appearing in the documents d1 to d5 are adopted as metadata, and the word frequency in the sentence is used as a weighting method. It is assumed that the method of using the weight value as it is is adopted.

  Specifically, for example, in the sentence d1, the frequency of each word of “Kyoto”, “Tofu”, “Onsen”, “Autumn leaves”, “USB”, “Software”, “Price” is 3, Suppose that it was 4, 1, 0, 0, 0, 1. In the sentence d2, the frequencies of the words “Kyoto”, “Tofu”, “Hot spring”, “Autumn leaves”, “USB”, “Software”, “Price” are 1, 0, 3, 3, 0 , 0, 1. In the sentence d3, the frequencies of the words “Kyoto”, “Tofu”, “Hot spring”, “Autumn leaves”, “USB”, “Software”, “Price” are 4, 1, 0, 0, 0. , 0, 2. In the sentence d4, the frequencies of the words “Kyoto”, “Tofu”, “Hot spring”, “Autumn leaves”, “USB”, “Software”, “Price” are 0, 1, 0, 4, 0. , 0, 0. In the sentence d5, the frequencies of the words “Kyoto”, “Tofu”, “Onsen”, “Autumn leaves”, “USB”, “Software”, “Price” are 0, 0, 0, 0, 2 , 1, 1.

  In this case, a weighted metadata matrix D as shown in FIG. 4 is generated as a result of the process of step S2. That is, as a result of the processing in step S2, each of the content vectors of the sentences d1 to d5 (content vectors weighted by frequency, which are so-called feature vectors) is changed to the first to fifth columns. A 7 × 5 metadata matrix D as a column component is generated.

  The content vector (“Kyoto”, “Tofu”, “Hot spring”, “Autumn leaves”, “USB”, “Software”, “Price”) of text d1 is (3,4,1,0,0, 0,1). The content vector of the sentence d2 is (1,0,3,3,0,0,1). The content vector of the sentence d3 is (4,1,0,0,0,0,2). The content vector of the sentence d4 is (0,1,0,4,0,0,0). The content vector of the sentence d5 is (0,0,0,0,2,1,1).

  Returning to FIG. 3, in step S <b> 3, the LSA operation unit 20 performs an LSA operation on the metadata matrix D appropriately weighted by the weighting processing unit 19 in the process of step S <b> 2.

Here, as the process of step S3, the first process and the third process of the LSA calculation are executed, and as a result, the approximate matrix _Dk appropriately dimensionally compressed is generated.

Specifically, in this case, when the process of step S3 is performed on the matrix D of FIG. 4, for example, an approximate matrix _Dk shown in FIG. 5 that is two-dimensionally compressed is generated.

In other words, as a result of the processing in step S3, the approximated matrix of 7 rows and 5 columns having the respective content vectors of the sentences d1 to d5 updated as follows as the respective column components of the first column to the fifth column. D _k will be generated.

  That is, the updated content vector of the sentence d1 is (3.6999, 2.6836, 0.7968, 0.1194, 0.0846, 0.0423, 1.6540). The updated content vector of the sentence d2 is (0.8301, 0.8297, 1.6489, 3.5394, 0.0168, 0.0084, 0.6448). The updated content vector of the sentence d3 is (3.2099, 2.3044, 0.5377, -0.2633, 0.0736, 0.0368, 1.4063). The updated content vector of the sentence d4 is (0.0886, 0.2855, 1.4478, 3.4166, -0.0001, -0.0001, 0.3057). The updated content vector of the sentence d5 is (0.2824, 0.2058, 0.0674, 0.0249, 0.0064, 0.0032, 0.1275).

Returning to FIG. 3, in step S4, the metadata extraction unit 21 calculates each feature difference of each metadata by using the approximate matrix _Dk calculated by the LSA calculation unit 15 in the process of step S3.

The feature difference indicates an index value of the importance level of metadata generated by using a difference (change) between the metadata matrix D and the approximate matrix _Dk .

For example, in the example of the approximate matrix _Dk in FIG. 5, the component indicated by two upward arrows such as “↑↑” has a weight value (component value) of 1 or more compared to the metadata matrix D in FIG. 4. Indicates an increase. Similarly, the component indicated by one upward arrow such as “↑” indicates that the weight value has increased by 0.5 or more compared to the metadata matrix D of FIG. The components indicated by two downward arrows such as “↓↓” indicate that the weight value has decreased by 1 or more compared to the metadata matrix D of FIG. Similarly, the component indicated by one downward arrow such as “↓” indicates that the weight value has decreased by 0.5 or more compared to the metadata matrix D of FIG.

Here, the meaning that the component value of the approximate matrix _Dk increases from that of the metadata matrix D will be described.

  That is, the time point of the metadata matrix D generated without considering the co-occurrence of the metadata across the respective contents although the importance of the predetermined metadata in the predetermined content is originally high. In this case, the degree of importance is regarded as low, and as a result, the corresponding component value of the metadata matrix D may be a low value.

In this case, when the approximate matrix D _k is generated, the original high importance of the metadata in the content is highlighted, and as a result, the corresponding component value of the approximate matrix D _k is updated to a high value. It is.

This is because the approximate matrix D _k is obtained as a result of reduction and recalculation of base components that are considered to be unimportant as a principal component in the concept space (having a small singular value). That is, the approximate matrix D _k is a matrix in which each component value is updated in consideration of the co-occurrence of metadata across each content.

  The above contents mean that the component value is increased by the LSA calculation.

Similarly, the meaning that the component value of the approximate matrix D _k is smaller than that of the metadata matrix D is as follows.

  That is, the time point of the metadata matrix D generated without considering the co-occurrence of the metadata across the respective contents although the importance of the predetermined metadata in the predetermined content is originally low. In this case, the degree of importance is considered high, and as a result, the corresponding component value of the metadata matrix D may be a high value.

In this case, when the approximate matrix D _k is generated, the inherent low importance of the metadata in the content is highlighted, and as a result, the corresponding component value of the approximate matrix D _k is updated to a low value. It is.

As described above, the difference (change) between the metadata matrix D and the approximate matrix _Dk is how to grasp the importance of the metadata before and after the co-occurrence of the metadata across the contents is considered. It can be said that the difference is expressed.

Therefore, by using the difference (change) between the metadata matrix D and the approximate matrix _Dk , it is possible to calculate an index value of the importance of the metadata, that is, a feature difference of the metadata.

In other words, the metadata feature difference calculation method is not particularly limited as long as it uses a difference (change) between the metadata matrix D and the approximated matrix _Dk, and various methods can be applied.

  For example, the feature difference of metadata can be calculated by the following first feature difference calculation method to third feature difference calculation method.

The first feature difference calculation method is a method for calculating a feature difference using the component values of the approximate matrix _Dk .

It is also because it can be said that using the difference (change) between the approximated matrix D _k metadata matrix D utilizing component value itself in the approximated matrix D _k.

Specifically, the predetermined metadata corresponds to one row in each of the metadata matrix D and the approximate matrix _Dk . For example, in the example of the metadata matrix D in FIG. 4 and the approximate matrix _{Dk in} FIG. 5 described above, the metadata (word) “Kyoto” corresponds to the first row. That is, each of the predetermined component values indicates a metadata weight value corresponding to the row for each content (sentence). For example, in the example of the metadata matrix D in FIG. 4 and the approximate matrix D _{k in} FIG. 5 described above, each component value in the first row is a meta-data “Kyoto” for each of the sentences d1 to d5. Refers to the weight value of data (word).

  Therefore, for example, when a metadata matrix D is generated from N pieces of metadata and M pieces of content, that is, when the metadata matrix D is a matrix of N rows and M columns, each of the N pieces of metadata Are sequentially set as metadata to be noted as a processing target (hereinafter referred to as “notable metadata”), and M component values in a row indicating the notable metadata, that is, M contents of the notable metadata. A method of calculating an average value or a maximum value of each weight value and using the calculation result as a feature difference of the target metadata is an example of a first feature difference calculation method.

The second feature difference calculation method is a method of calculating a feature difference using each of the difference values between the values of the respective components of the approximate matrix D _k and the values of the corresponding components of the metadata matrix D. is there.

Specifically, for example, when the metadata matrix D is a matrix of N rows and M columns, each of the N pieces of metadata is sequentially set as attention metadata, and a row indicating attention metadata in the approximated matrix _Dk. Each of the M component values and the corresponding component value of the metadata matrix D are calculated, and an average value or a maximum value of the calculated M difference values is calculated. A technique of setting the calculation result as a feature difference of the target metadata is an example of a second feature difference calculation technique.

If the predetermined component value is increased by LSA computing, i.e., in a given component, if it is a value larger metadata matrix D than the approximated matrix D _k, for that component, approximated matrix D _k Naturally, the difference value between the metadata matrix D is a positive value.

  Considering the above contents and the above-mentioned meaning that the component value becomes larger by the LSA calculation, the feature difference of the metadata of interest calculated by the second feature difference calculation method eventually becomes a positive value. Metadata indicates that the metadata is determined to be important as a result of considering the co-occurrence of metadata across contents. To be precise, being positive means that the original high importance of the metadata of interest has been highlighted.

  In addition, when the feature difference of the target metadata calculated by the second feature difference calculation method becomes a negative value for the reason opposite to the reason of becoming a positive value, the metadata spans each content. As a result of considering the co-occurrence of metadata, it means that the importance is determined to be low. To be precise, becoming negative means that the original low importance of the metadata of interest has been highlighted.

Specifically, for example, the calculation result of the feature difference calculated by the second feature difference calculation method using the approximate matrix _Dk of FIG. 5 is shown in FIG. More precisely, the words “Kyoto”, “Tofu”, “Onsen”, “Autumn leaves”, “USB”, “Software”, “Price” are set as attention metadata in order, and the approximation of FIG. Five component values of the row indicating the target metadata in the matrix D _k , that is, each of the weight values of the target metadata in the sentences d1 to d5, and the corresponding component value in the metadata matrix D of FIG. FIG. 6 shows the calculation result when each of the difference values is calculated and the average value of these five difference values is calculated as the feature difference of the metadata of interest.

  Specifically, as shown in FIG. 6, the feature difference value of “Kyoto” is 0.022. In addition, the characteristic difference values of “tofu”, “hot spring”, “colored leaves”, “USB”, “software”, “price” are 0.0618, 0.0997, -0.326, -0.3638, -0.1819, -0.1723, respectively. Become each.

  Therefore, each of “Kyoto”, “Tofu”, and “Hot Spring” was determined to have high importance as a result of considering the co-occurrence of words across sentences d1 to d5. It can be said that high importance has been highlighted.

  On the other hand, “Autumn Leaves”, “USB”, “Software”, and “Price” are determined to be low in importance as a result of considering the co-occurrence of words across sentences d1 to d5. To be precise, it can be said that the original low importance is highlighted.

  More specifically, the following can be understood from the feature difference value of each metadata in FIG. That is, the importance of words that are not related to other words except for each other, such as “USB” and “software” that appear only in the sentence d5 that is not related to other documents, is very low (the weight is large). You can see). It can also be seen that the importance of general words that tend to appear in any sentence, such as “price”, is reduced (the weight is reduced). On the other hand, it can be seen that the importance of words such as “hot spring” and “tofu” that characterize the document and that have a plurality of similar documents becomes high (weight increases).

  The second feature difference calculation method has been described above. Next, a third feature difference calculation method will be described.

The third feature difference calculation method is a method of calculating a feature difference using each of the division values obtained by dividing the values of the respective components of the approximate matrix D _k by the values of the corresponding components of the metadata matrix D. It is.

Specifically, for example, when the metadata matrix D is a matrix of N rows and M columns, each of the N pieces of metadata is sequentially set as attention metadata, and a row indicating attention metadata in the approximated matrix _Dk. Each of the M component values is divided by the corresponding component value in the metadata matrix D, and the average or maximum value of the calculated M divided values is calculated. A technique of setting the calculation result as a feature difference of the target metadata is an example of a third feature difference calculation technique.

If the predetermined component value is increased by LSA computing, i.e., in a given component, if it is a value larger metadata matrix D than the approximated matrix D _k, for that component, approximated matrix D _k The division value of the metadata matrix D with respect to is naturally larger than 1.

  Considering the above contents and the above-mentioned meaning that the component value is increased by the LSA calculation, the feature difference of the target metadata calculated by the third feature difference calculation method is eventually larger than 1. Metadata indicates that the metadata is determined to be important as a result of considering the co-occurrence of metadata across contents. To be precise, being larger than 1 means that the original high importance of the metadata of interest has been highlighted.

  In addition, the feature difference of the target metadata calculated by the second feature difference calculation method is smaller than 1 for the reason opposite to the reason that the target metadata is greater than 1. The target metadata spans each content. As a result of considering the co-occurrence of metadata, it means that the importance is determined to be low. To be precise, being smaller than 1 means that the original low importance of the metadata of interest has been highlighted.

  As described above, the first feature difference calculation method to the third feature difference calculation method have been described as examples of the metadata feature difference calculation method in step S4 of FIG.

  In this manner, when each feature difference of each metadata is calculated by the process of step S4, the process proceeds to step S5.

  In step S5, the metadata extraction unit 21 determines whether or not the metadata feature difference is equal to or less than a threshold value.

  If all of the feature differences of each metadata exceed the threshold value, it is determined as NO in step S5, and the process ends.

  On the other hand, if at least one feature difference equal to or less than the threshold exists among the feature differences of each metadata, it is determined as YES in Step S5, and the process proceeds to Step S6.

  In step S6, the metadata extraction unit 21 registers and presents unnecessary metadata. Specifically, in step S6, the metadata extraction unit 21 specifies that metadata having a feature difference equal to or less than a threshold among the feature differences of each metadata is unnecessary metadata, and stores the unnecessary metadata in the metadata. Extract from part 16. The metadata extraction unit 21 registers (stores) the extracted unnecessary metadata in the user dictionary storage unit 13 or the general dictionary storage unit 14 or presents it to the user via the user interface unit 11. Thereby, the “unnecessary metadata extraction process considering the co-occurrence relationship” is completed.

  As described above, the threshold used in the process of step S5 is a value that is compared with the feature difference of metadata in order to determine whether or not each content is classified as unnecessary metadata. That is, metadata having a feature difference exceeding a threshold is highly important metadata that is not classified as unnecessary metadata. On the other hand, metadata having a feature difference that is less than the threshold is metadata with low importance that is classified as unnecessary metadata.

  Therefore, this threshold value is often different depending on the feature difference calculation method employed in the process of step S4.

  For example, when the above-described second feature difference calculation method using a difference value is adopted, it is preferable to set a value less than 0 as the threshold value, for example. Specifically, for example, when −0.1 is set as the threshold value, “USB”, “software”, and “price” are extracted as unnecessary metadata in the example of FIG. 6 described above.

  On the other hand, for example, when the above-described third feature difference calculation method using a division value is adopted, it is preferable to set a value less than 1, for example, as the threshold value.

  2 to 6, the information processing system or information processing apparatus according to the first embodiment, that is, the information processing system or information processing apparatus that executes “unnecessary metadata extraction processing considering the co-occurrence relationship”. Explained.

In the first embodiment, by utilizing the approximate matrix D _k or the difference between the approximate matrix D _k and the original metadata matrix D, the relationship (co-occurrence) between the metadata at the potential semantic level is used. The weight is taken into consideration. That is, an index value of importance that takes into account the co-occurrence relationship such as a feature difference is obtained.

  Therefore, by using the importance index value (weight value) considering such a co-occurrence relationship, it seems that it is related to metadata that does not seem to be related to others, or is seemingly related to others. It is possible to discover metadata that has low reliability and make selections based on it.

  In other words, it is possible to prevent metadata that is seemingly unrelated to others but that is originally highly important metadata from being misclassified as unnecessary metadata. Also, metadata that seems to be related to others at first glance and is actually less relevant, that is, metadata that seems to be highly important at first glance but is actually less important can be reliably classified as unnecessary metadata. It becomes possible.

  (Second Embodiment)

  Next, a second embodiment will be described.

In the conventional content recommendation, the co-occurrence relationship of metadata is not considered, but the weight in the metadata matrix D by TF / IDF, or the approximate matrix D obtained as a result of dimension compression of the metadata matrix D by LSA. The weight in _k was used, and there was a problem that any method could only recommend content similar to what was known (experienced or highly rated by the user).

  In order to solve this problem, the present inventor has invented the above-described second process, that is, the “recommendation process considering the co-occurrence relationship”.

This second processing uses the approximate matrix D _k generated by the LSA or the feature difference of the metadata described in the first embodiment. As described above, the approximate matrix D _k is a matrix generated in consideration of the co-occurrence relationship of metadata, and the feature difference of the metadata has an importance level considering the co-occurrence relationship for the metadata. This is because it is an index value.

  The outline of the second process will be described below.

The information processing system or information processing apparatus of the second embodiment (hereinafter simply referred to as “device” in the description of the outline of the second processing) has a feature difference or approximation when focusing on a certain content (column vector). One or more metadata used for content recommendation are extracted based on the component value of the matrix _Dk .

  Specifically, as described above, the metadata having a large feature difference is determined to be important in consideration of the co-occurrence relationship with other metadata, although the weight is not so large in the original metadata matrix D. Metadata (hereinafter referred to as important metadata). Therefore, the important metadata mentioned here is considered to be highly emergent metadata that the user has not noticed before.

  Therefore, for example, the apparatus can extract the top few metadata having a large feature difference as important metadata.

Also, it can be said that metadata corresponding to a large component value in the approximate matrix D _k is also important metadata.

Therefore, the apparatus can extract, for example, metadata corresponding to the top several component values of the approximate matrix D _k as important metadata.

Alternatively, the device can extract the important metadata based on the feature difference and also extract the important metadata based on the component values of the approximate matrix _Dk . That is, as the one or more important metadata used for content recommendation, only the important metadata extracted based on the feature difference may be used, or the important metadata extracted based on the component values of the approximate matrix D _k. Or the important metadata extracted based on the feature difference and the important metadata extracted based on the component values of the approximate matrix D _k may be used in combination.

  After that, the apparatus recommends the one or more important metadata itself extracted in this way as information that is a trigger for the user to select content. Alternatively, the apparatus regards a metadata group including one or more important metadata extracted in this way as one content (column vector), and the metadata group (column vector) and other content (column vector). And recommends other content based on the result of the matching process.

  The outline of the second process, that is, the “recommendation process considering the co-occurrence relationship” has been described above.

  Next, the information processing system or information processing apparatus according to the second embodiment, that is, the information processing system or information processing apparatus that executes the “recommendation process considering the co-occurrence relationship” will be described with reference to FIGS. To do.

  FIG. 7 illustrates a functional configuration example of the information processing system or the information processing apparatus according to the second embodiment.

  In other words, blocks necessary for executing the “recommendation process considering the co-occurrence relationship” are extracted from all the blocks of the user interface unit 11 to the content recommendation unit 23 in FIG. FIG. 7 is a diagram arranged in accordance with the information flow at the time of the “recommendation process taking into account”. Accordingly, the description of each block shown in FIG. 7 has been described above with reference to FIG.

  Although omitted in the example of FIG. 7, the information transmission unit 24 of FIG. 1 is actually arranged in each arrow connecting the two blocks, that is, between the two blocks. become.

  FIG. 8 is a flowchart for explaining an example of the “recommendation process considering the co-occurrence relationship”. Therefore, an example of “recommendation processing considering co-occurrence relations” will be described below with reference to the flowchart of FIG.

  Each of steps S21 to S23 in FIG. 8 is basically the same processing as each of steps S1 to S3 in FIG. 3 described above. Therefore, the description of the processing in steps S21 to S23 is omitted.

However, the more the content (content vector) irrelevant to the user experience is included in the metadata matrix D generated in the process of step S21, the more the approximate matrix D _k obtained as a result of the process of step S23 is. This is a matrix in which the bias of the co-occurrence relationship of user-specific metadata is thinned, and the matrix in which the co-occurrence relationship in a general sense is taken into consideration. Therefore, the metadata extracted in the process of step S24 described later based on each component value of the approximate matrix _Dk or the feature difference obtained from the approximate matrix _{Dk is} generated for the user. Note that this is a degraded metadata, so be careful. That is, when it is desired to extract metadata that has high emergence for the user, it is preferable to include as much content (content vector) that the user has experienced in the metadata matrix D generated in the process of step S21 as much as possible. .

When the approximate matrix D _k is generated by the LSA calculation unit 20 in the process of step S23, the process proceeds to step S24.

  In step S24, the LSA calculation unit 20 determines whether or not to use the feature difference in the process of step S26 described later executed by the metadata extraction unit 21.

  If it is determined in step S24 that the feature difference is used, the LSA calculation unit 20 calculates the feature difference of each metadata in step S25. Note that the process of step S25 is basically the same process as the process of step S4 of FIG. 3 described above. Therefore, the details of the process in step S25 are omitted.

Thereafter, when the approximate matrix D _k and the feature difference of each metadata are supplied from the LSA calculation unit 20 to the metadata extraction unit 21, the process proceeds to step S26.

On the other hand, when it is determined in step S24 that the feature difference is not used, only the approximate matrix _Dk is supplied from the LSA calculation unit 20 to the metadata extraction unit 21, and the process proceeds to step S26.

In step S _< b> 26, the metadata extraction unit 21 sets metadata to be used for recommendation using at least one of the component value of the approximate matrix D _k and the feature difference of each metadata, that is, important metadata as 1. As described above, one or more specified metadata is extracted from the metadata storage unit 16.

  The extraction method (specific method) of important metadata in step S26 is not particularly limited. For example, the following extraction method can be employed.

That is, for example, in all column components of the approximate matrix D _k , that is, an average vector of all content vectors or a specific content vector specified by the user, metadata corresponding to the highest component value (or an arbitrary number from the higher one) It is possible to apply an extraction method that extracts (metadata) as important metadata, or an extraction method that uses the component values of the approximate matrix _Dk .

  Also, for example, extraction of metadata having the highest feature difference (or an arbitrary number of metadata from the highest) as important metadata, that is, extracting metadata with an increased weight value as important metadata To summarize the technique, it is possible to apply an extraction technique that uses feature differences.

Specifically, for example, it is assumed that the approximate matrix _{Dk of} FIG. 5 is generated from the metadata matrix D of FIG. 4 described in the first embodiment in the processes of steps S21 to S23 described above. Further, in the process of step S25, the feature difference of each metadata in FIG. 6 is obtained by the above-described second feature difference calculation method using the difference value between the approximate matrix _Dk in FIG. 5 and the metadata matrix D in FIG. Is calculated.

  In this case, if metadata having a feature difference of 0.05 or more is extracted as important metadata in the process of step S26, "tofu" and "hot spring" are extracted.

  When one or more important metadata extracted by the metadata extraction unit 21 is supplied to the content recommendation unit 23, the process proceeds to step S27.

  In step S27, the content recommendation unit 23 determines whether or not to recommend content.

  If it is determined in step S27 that no content is recommended, the process proceeds to step S28.

  In step S28, the content recommendation unit 23 presents one or more important metadata extracted by the metadata extraction unit 21 in the process of step S26 to the user via the user interface unit 11.

  Thereby, the “recommendation process considering the co-occurrence relationship” is completed.

  On the other hand, if it is determined in step S27 that the content is recommended, the process proceeds to step S29. To be exact, when it is determined in step S27 that content recommendation is to be performed, the content recommendation unit 23 supplies one or more important metadata extracted by the metadata extraction unit 21 in the process of step S26 to the vector calculation unit 22. By supplying and requesting the matching process, the process proceeds to step S29.

  In step S29, the vector calculation unit 22 performs content matching processing using a metadata group including one or more important metadata extracted by the metadata extraction unit 21 in step S26. That is, in step S29, the vector calculation unit 22 regards a metadata group including one or more important metadata extracted by the metadata extraction unit 21 in the process of step S26 as one content (content vector), and The degree of similarity with other content (content vector) stored in the content storage unit 15 is calculated, and the content with the highest similarity (or any number of content from the highest) is selected and supplied to the content recommendation unit 23 To do.

  In step S28, the content recommendation unit 23 recommends one or more contents selected by the vector calculation unit 22 in step S29. That is, in step S28, the content recommendation unit 23 presents one or more pieces of metadata (or related information) of the content to the user via the user interface unit 11.

  Thereby, the “recommendation process considering the co-occurrence relationship” is completed.

  The information processing system or information processing apparatus according to the second embodiment, that is, the information processing system or information processing apparatus that executes the “recommendation process considering the co-occurrence relationship” has been described above with reference to FIGS. 7 and 8. .

In the second embodiment, an approximate matrix D _k is obtained. By using the approximate matrix D _k and the difference between the approximate matrix D _k and the original metadata matrix D, etc., at a potential semantic level. Are weighted in consideration of the relationship (co-occurrence relationship) between the metadata. That is, an approximate matrix D _k that takes into account the co-occurrence relationship of metadata is obtained, and a feature difference that is an index value of importance taking into account the co-occurrence relationship of metadata is obtained.

Therefore, by using the component values of the approximate matrix D _k considering such a co-occurrence relationship and the importance index values (weight values) considering the co-occurrence relationship, a meta that has no relation to others at first glance is used. It is possible to find metadata and metadata that seem to be related to others at first glance and actually have low relevance, and make selections based on it.

  In other words, it is metadata that seems to be unrelated to others at first glance. However, as described above, metadata with high importance is, as described above, highly emergent metadata that has not been noticed by the user, that is, important metadata. It is thought that. Therefore, the content recommended based on such important metadata is also considered to be highly emergent content that the user has not noticed before.

  The information processing system or information processing apparatus according to the first embodiment or the second embodiment described above is applied to attribute (metadata) selection processing called feature selection or the like in fields such as data mining and document classification. You can also. That is, an attribute (metadata) selection process considering the co-occurrence relationship of metadata can be easily realized.

  (Third embodiment)

  Next, a third embodiment will be described.

  As described above in [Problems to be Solved by the Invention], a user has received high evaluation as a method for generating a user preference vector (UPV) of a content recommendation system based on the vector space method. In many cases, a generation method for generating a UPV based on an average of content vectors of content groups is employed. The UPV generated by such a generation method is a vector that imitates the user's various preferences, and even if content recommendation is performed using such UPV, it is difficult to make a wide range of recommendations was there. In addition, even if a highly evaluated content group is clustered into a plurality of groups to produce a variety, there is a problem that it is difficult to recommend content that the user has never experienced.

  In order to solve this problem, the present inventor invented the above-described third processing, that is, “recommendation processing using a difference between clustered UPV groups”.

  The outline of the third process will be described below.

  The information processing system or information processing apparatus of the third embodiment (hereinafter simply referred to as “device” in the description of the outline of the third processing) is a content vector that is highly evaluated by the user in the metadata space or the concept space. Are clustered into a plurality of groups using an arbitrary algorithm.

  For each group, the apparatus obtains a representative vector (hereinafter referred to as a representative vector) by averaging one or more content vectors belonging to the corresponding group, and further obtains a difference between representative vectors of each group. Then, a vector having the difference as a component (hereinafter referred to as difference UPV) is generated.

  That is, the vector group composed of the representative vectors of each group in the third embodiment is a clustered conventional UPV group, and the difference UPV is a vector generated by the difference of the clustered conventional UPV group. .

  The apparatus performs content matching processing using the differential UPV, and recommends content based on the result of the matching processing.

  What should be noted here is that the difference UPV is a vector that represents a preference that cannot be calculated as an average of content vectors (it cannot be calculated). Therefore, by using the differential UPV, it is possible to recommend content that the user has not noticed before.

  The outline of the third process, that is, the “recommendation process using the difference between clustered UPV groups” has been described above.

  Next, referring to FIG. 9 and FIG. 10, the information processing system or information processing apparatus of the third embodiment, that is, the information processing system or information that executes “recommendation processing using the difference between clustered UPV groups” The processing apparatus will be described.

  FIG. 9 illustrates a functional configuration example of the information processing system or the information processing apparatus according to the third embodiment.

  In other words, blocks necessary for executing the “recommendation process using the difference between clustered UPV groups” are extracted from all the blocks of the user interface unit 11 to the content recommendation unit 23 in FIG. FIG. 9 is a diagram arranged according to the information flow when executing the “recommendation process using the difference between clustered UPV groups”. Therefore, the description of each block shown in FIG. 9 has been described above with reference to FIG.

  Although omitted in the example of FIG. 9, the information transmission unit 24 of FIG. 1 is actually arranged in each arrow connecting the two blocks, that is, between the two blocks. become.

  FIG. 10 is a flowchart for explaining an example of the “recommendation process using a difference between clustered UPV groups”. Therefore, an example of “extraction processing using a difference between clustered UPV groups” will be described below with reference to the flowchart of FIG.

  Each of steps S41 and S42 in FIG. 10 is basically the same processing as each of steps S1 and S2 in FIG. 3 described above. Therefore, the description of the processes in steps S41 and S42 is omitted.

  For example, it is assumed that a matrix A in which content vectors with high user evaluation are collected is generated as the metadata matrix D weighted in the processes of steps S41 and S42. Hereinafter, each column component of A, that is, each content vector is described as ai (i = 0, 1,..., M−1). That is, the matrix A is expressed by the following equation (3).

  A = (a0, a1,..., Am-1) (3)

  In this case, in step S43, the LSA calculation unit 20 performs an LSA calculation on the metadata matrix A appropriately weighted by the weighting processing unit 19 in the process of step S42.

  However, in the process of step S43 of the third embodiment, the first process and the second process of the LSA calculation are executed.

  Specifically, for example, in the LSA calculation in step S43, the matrix A is decomposed into each of the three component matrices U, Σ, and V by the singular value decomposition expressed by the above-described equation (1).

Next, in the LSA calculation in step S43, the component matrix U is compressed to k dimensions, and as a result, a projection matrix U _k is obtained. That is, the projection matrix U _k refers to a matrix in which only k column components (column vectors) from the component matrix U having the largest singular value are left and the other components are zero.

Next, in the LSA calculation of step S43, the matrix A is projected onto the concept space by the projection matrix U _k . Note that a matrix obtained as a result is described as a matrix B, for example. In this case, the fact that the matrix A is projected onto the concept space by the projection matrix U _k indicates that an operation according to the following equation (4) has been performed. In Equation (4), the matrix U _k ^T represents a transposed matrix of the projection matrix U _k .

B = U _k ^T A (4)

  Each column component (column vector) of the matrix B is described as bi (i = 0, 1,..., M−1). That is, the matrix B is expressed by the following equation (5).

  B = (b0, b1, ..., bm-1) (5)

  This column vector bi is a content vector compressed in the k dimension, that is, a content vector projected onto the concept space.

  That is, in the process of step S43, each content vector bi projected into the concept space is obtained. Hereinafter, an aggregate of each content vector bi projected onto the concept space, that is, the matrix B is referred to as a content vector group projected onto the concept space.

  Therefore, in step S44, the vector calculation unit 22 performs clustering of content vector groups projected onto the concept space by the processing of the LSA calculation unit 20 in step S43. That is, in step S44, the vector calculation unit 22 classifies each content vector bi projected onto the concept space into one of an arbitrary number of arbitrary types of clusters using an arbitrary algorithm.

  It can be said that the vector calculation unit 22 that executes the process of step S44 is the clustering unit 22. Therefore, the vector calculation unit 22 shown below the LSA calculation unit 20 in FIG. 9 is shown in parentheses with the clustering unit 22.

  Specifically, for example, it is now assumed that each content vector bi projected onto the concept space is classified into one of s clusters in step S44.

  Next, in step S45, the vector calculation unit 22 generates a representative vector (UPV). That is, in this case, in step S45, the vector calculation unit 22 calculates, for each of the s clusters, an average vector of one or more content vectors bi belonging to the corresponding cluster, that is, each component of the one or more content vectors bi. Is generated as a component value, and the average vector is used as a representative vector (UPV).

  Hereinafter, this representative vector is described as cj ′ (j = 0, 1,..., S−1).

  In step S46, the vector calculation unit 22 generates a difference UPV that is a difference between representative vectors (UPV). That is, in step S46, the vector calculation unit 22 generates one difference vector by obtaining a difference between two arbitrary combinations of the representative vectors cj ′ of the s clusters.

  The total number of such combinations of two clusters differs depending on the number of clusters s. However, when the number of clusters s is 3 or more, it is naturally plural. Accordingly, in this case, if the difference UPVs are generated for all the combinations, a plurality of difference UPVs are generated.

  Specifically, for example, in the present case, in the process of step S46, the right side of the following equation (6) is calculated, and each of the vectors d'p, q is generated as each difference UPV. In Equation (6), p, q = 0, 1,..., S−1. However, p ≠ q.

  d’ p, q = c’p−c’q (6)

  Note that the combinations for generating the differential UPV need not be all combinations, and may be an arbitrary number of arbitrary combinations. In any case, one or more differential UPVs are generated in the process of step S46. Therefore, hereinafter, one or more differential UPVs are defined as a differential UPV group. That is, the difference UPV group is generated in the process of step S46.

  Further, as the processing of step S46, the vector calculation unit 22 further follows a predetermined rule such as the descending order of the value of the first principal component (vector basis value paired with the highest singular value by singular value decomposition) in the concept space. Each differential UPV belonging to the differential UPV group can also be ordered.

  When generating the difference UPV group, the vector calculation unit 22 notifies the content recommendation unit 23 to that effect. Thereafter, when a request for matching processing is notified from the content recommendation unit 23 to the vector calculation unit 22, the process proceeds to step S47.

  In step S47, the vector calculation unit 22 performs content matching processing using the differential UPV group generated in step S46.

  That is, in step S47, the vector calculation unit 22 calculates the similarity between each difference UPV belonging to the difference UPV group and the other content (content vector) stored in the content storage unit 15, and the most similar A content with a higher degree (or an arbitrary number of contents from the higher one) is selected and supplied to the content recommendation unit 23.

  Specifically, for example, in this case, each vector d'p, q (p, q = 0,1, ..., s-1, where p ≠ q) belongs to the differential UPV group. In the process of step S47, the similarity between the corresponding vector d'p, q and the new content vector is calculated for all p, q (or for the top several if ranked).

  In addition, it can be said that the vector calculation part 22 which performs this process of step S47 is the matching part 22 with respect to the vector calculation part 22 which performs the process of step S44. Therefore, the vector calculation unit 22 shown on the right side of the content recommendation unit 23 in FIG. 9 is shown in parentheses with the matching unit 22.

  In step S48, the content recommendation unit 23 recommends one or more contents selected by the vector calculation unit 22 in the process of step S47. That is, in step S48, the content recommendation unit 23 presents one or more pieces of metadata (or related information) of the content to the user via the user interface unit 11.

  Thus, the “recommendation process using the difference between clustered UPV groups” is completed.

  As described above, with reference to FIGS. 9 and 10, the information processing system or information processing apparatus according to the third embodiment, that is, the information processing system or information processing that executes the “recommendation process using the difference between clustered UPV groups”. The apparatus has been described.

  In the third embodiment, the following effects can be obtained. That is, conventionally, as described above, the UPV is generated from the average of content vectors having a high user evaluation. Therefore, content having a high degree of similarity with UPV inevitably resembles the content experienced by the user, and there is a problem that variations in content recommendation are narrow. On the other hand, in the third embodiment, content recommendation is performed based on the result of the matching process using the differential UPV, so that the content of the content that has never experienced the user and reflects the user's preference to some extent It is possible to produce an effect that recommendation can be made.

  Hereinafter, this effect of the third embodiment will be further described. In addition, in order to make an understanding easy, it demonstrates, referring suitably each step shown by the flowchart of FIG. 10 mentioned above.

  In the metadata space before projection, that is, before the processing in step S43, if the metadata matrix D is generated using, for example, the word frequency in the document, its column component, that is, the content vector Negative vector elements (negative component values, hereinafter referred to as negative elements) have no meaning.

  Therefore, in the metadata space, content vector groups are clustered, representative vectors (UPV) of each cluster are generated, and even if the difference between representative vectors UPV is taken, matching between the resulting vector and content is performed. In processing, negative elements cannot be used as information

  On the other hand, after the process of step S43, that is, in the concept space obtained as a result of projecting the metadata space by singular value decomposition, as described above, each content vector has a negative element. Become.

  Therefore, in the concept space, when the difference UPV obtained as a result of steps S44 to S46 described above is used in the matching process in step S47, that is, the content vector group projected on the concept space is clustered, and each cluster When the representative vector (UPV) is generated and the difference UPV generated by taking the difference between the representative vectors UPV is used for the matching process, all the elements including the negative element are valid.

  Specifically, for example, in the process of step S44, clustering is performed according to the user's preference in the concept space, and the cluster representative vector c1 indicating the first preference has a high weight on the concept basis e1, e2, e3. The representative vector c2 of the cluster indicating the second preference different from the first preference is assumed to have a high weight on the conceptual basis of e2, e3, e4. For simplicity of explanation, the weight values (component values) of e1 to e4 are all positive values.

  Note that the concept basis refers to a basis that extends the concept space. Specifically, for example, each column component (column vector) of the component matrix U when the metadata matrix D is subjected to singular value decomposition according to the above-described equation (1). Point to.

  In this case, in the vector (c1-c2) which is the difference UPV between the representative vector c1 and the representative vector c2, the positive high weight value of the concept base e1 and the negative high weight value of the concept base e4 remain. That is, in the concept base e2 and the concept base e3, as a result of taking the difference between the high weight value and the high weight value, the weight values of the two cancel each other, and the absolute value of the weight value is the concept base e1, The value is much lower than the absolute value of the weight value of e4.

  Accordingly, in step S47, such content that matches the difference UPV has a high weight on the metadata projected onto the concept base e1 and a high weight on the metadata projected in the negative direction with respect to the concept base e4. It can be said that there is something. The metadata projected in the negative direction on the concept base e4 may not be added to the content experienced by the user, even if there is something related to the metadata projected in the positive direction of the concept bases e1 to e4 There is sex. For this reason, by adding the metadata projected in the negative direction to the concept base e4 to the target of the matching process, it is possible to perform content recommendation that can trigger a new interest of the user.

  (Fourth embodiment)

  Next, a fourth embodiment will be described.

  Conventionally, content recommendation using user evaluation values has been performed. For example, P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl. “GroupLens: Open Architecture for Collaborative Filtering of Newnews.” Conference on Computer Supported Cooperative Work, pp. 175-186, 1994. Discloses a content recommendation technique using collaborative filtering and user evaluation values. Japanese Patent Laid-Open No. 2002-269143 discloses a content recommendation method using LSA and a user evaluation value.

  However, these methods simply use the similarity of evaluation between different users, and the temporal change of evaluation for content having the same tendency within one user and the grasped contents Is not considered. Therefore, there is a problem that the content recommended by such a method is not necessarily suitable for the current user's preference.

  Therefore, in order to solve this problem, the present inventors have invented the above-described fourth processing, that is, “content re-evaluation processing by LSA”.

  The outline of the fourth process will be described below.

For example, now the content (new content) experienced by the user has increased, and accordingly, the information processing system or information processing apparatus of the fourth embodiment (hereinafter simply referred to as “device” in the outline of the fourth process). Assume that the metadata matrix D is updated by adding the content vector of the new content to the original metadata matrix D, and an approximate matrix D _k of the updated metadata matrix D is generated. That is, it is assumed that the approximate matrix _Dk has been updated.

In this case, the component value of the content vector (column component) included in the approximate matrix D _k before the update changes in the approximate matrix D _k after the update.

  Therefore, in the fourth embodiment, a content vector having a user evaluation value as a base in addition to metadata is used, and a metadata matrix D is generated from such content vector.

Thereafter, when the content experienced by the user (new content) increases and the user's evaluation value for the new content is also input, the apparatus vectorizes the new content based on the metadata and the user's evaluation value. Thereby, a content vector of the new content is generated. Then, the device updates the metadata matrix D by adding the content vector of the new content to the original metadata matrix D, and generates an approximate matrix _Dk of the updated metadata matrix D. That is, the approximate matrix _Dk is updated.

In this case, as described above, the evaluation value of the existing content similar to the new content (updated approximate matrix D) is determined by the evaluation value of the content vector of the new content (corresponding component value of the updated metadata matrix B). The corresponding evaluation value of _k ) will also change.

In other words, it can be said that the apparatus is performing reevaluation (update of the evaluation value) of the existing content by updating the approximate matrix D _k so as to include the content vector of the new content.

  As a result of such re-evaluation of existing content, the evaluation value of content that has not reached the reference value of the user recommendation target may reach the reference value after LSA execution. In such a case, the apparatus can recommend to the user content having an evaluation value that has reached the reference value after execution of LSA. In other words, the device can make recommendations according to the current user's preference from contents that were not recommended in the past, that is, contents that were not recommended in the past and were discarded. It becomes. In other words, it is possible to cope with a change in preference over time.

  The outline of the fourth process, that is, the “content re-evaluation process by LSA” has been described above.

  Next, an information processing system or information processing apparatus according to the fourth embodiment, that is, an information processing system or information processing apparatus that executes “content reevaluation processing by LSA” will be described with reference to FIGS. 11 and 12. .

  FIG. 11 illustrates a functional configuration example of the information processing system or the information processing apparatus according to the fourth embodiment.

  In other words, blocks necessary for executing the “content re-evaluation process by LSA” are extracted from all the blocks of the user interface unit 11 to the content recommendation unit 23 of FIG. FIG. 11 is a diagram arranged in accordance with the information flow at the time of executing the “re-evaluation process”. Therefore, the description of each block shown in FIG. 11 has been described above with reference to FIG.

  Although omitted in the example of FIG. 11, the information transmission unit 24 of FIG. 1 is actually arranged in each arrow connecting the two blocks, that is, between the two blocks. become.

  FIG. 12 is a flowchart for explaining an example of “content re-evaluation processing by LSA”. Accordingly, an example of “content reevaluation processing by LSA” will be described below with reference to the flowchart of FIG.

  In order to facilitate understanding of the “content re-evaluation process by LSA”, the following description will be given with reference to FIGS. 13 to 15 as appropriate. That is, FIGS. 13 to 15 show specific examples of processing results of “unnecessary metadata extraction processing considering co-occurrence relationships”.

  Here, for example, as shown in FIGS. 13 to 15, it is assumed that a piece of music is a processing target and the feature amount of the piece of music is adopted as metadata. Specifically, for example, as shown in FIGS. 13 to 15, five feature amounts such as “tempo”, “brightness”, “slow / slow”, “volume”, and “sound density” are employed. To do. Further, it is assumed that “evaluation” which is a user evaluation value for music is added to the base of the content vector in addition to these five feature amounts. That is, as shown in FIG. 13 to FIG. 15, here, the content vector is (“tempo”, “brightness”, “slow / slow”, “volume”, “sound density”, “evaluation”). Is a vector of the form.

In addition, the “content re-evaluation process by LSA” in which the four music pieces t1 to t4 are processed is performed in the past. At this time, the metadata matrix D0 of FIG. 13 is generated, and the metadata matrix D0 is also generated. There as a result of being compressed in two dimensions by LSA computing, and approximated matrix D0 _k in FIG. 14 is generated.

  As illustrated in FIG. 13, the metadata matrix D0 is a 6 × 4 matrix in which each content vector of the music pieces t1 to t4 is each column component of the first column to the fourth column. The content vector of the music t1 is (3,4,1,1,1,2). The content vector of the music piece t2 is (1,1,3,3,1,3). The content vector of the music t3 is (1,1,1,4,3,4). The content vector of the music t4 is (1,1,3,1,2,1).

Further, as shown in FIG. 14, the approximate matrix D0 _k represents the content vectors of the music pieces t1 to t4 updated as follows and the respective column components of the first to fourth columns. This is a 6 × 4 matrix. The updated content vector of the music t1 is (2.9829,3.9135,1,1460,0.9474,1.3666,1.8780). The updated content vector of the music piece t2 is (1.0413, 1.0535, 1.8432, 3.2809, 1.1293, 3.2931). The updated content vector of the musical piece t3 is (0.9531, 0.8869, 2.0439, 3, 7325, 1.1950, 3.6664). The updated content vector of the music t4 is (1.0503, 1.2953, 0.7850, 1.1136, 0.6536, 1.3586).

  Thereafter, it is assumed that the user views the new music t5 and evaluates the new music t5 using the user interface unit 11 of FIG. In this case, the ID and evaluation value of the new song t5 are stored in the user profile storage unit 12, and the metadata of the new song t5, that is, “tempo”, “brightness”, “slow”, “ The “volume” and “sound density” are stored in the metadata storage unit 16.

  As a result, it is assumed that the “content re-evaluation process by LSA” in FIG. 12 is started.

  In this case, in steps S61 and S62, processing similar to that in steps S1 and S2 in FIG. 3 is executed, and for example, the matrix matrix D as shown in FIG.

  Specifically, for example, (4, 2, 1, 1, 1, 5) is generated as the content vector of the music t5, and the content vector of the music t5 is added to the metadata matrix D0 in FIG. Metadata matrix D is generated.

  In this way, by the processing in steps S61 and S62, a 6-row and 5-column matrix in which each content vector of the music pieces t1 to t5 is a column component in each of the first to fifth columns is a metadata matrix. Generated as D. When the metadata matrix D is supplied from the weighting processing unit 19 to the LSA calculation unit 20, the process proceeds to step S63.

  Returning to FIG. 12, in step S63, the LSA operation unit 20 performs the LSA operation on the metadata matrix D of FIG.

  In this case, as the process of step S63, the first process and the third process of the LSA calculation are executed, and as a result, for example, the approximate matrix Dk shown in FIG. 16 compressed in two dimensions is generated. Is done.

  In other words, in this case, as a result of the process of step S63, the content vectors of the music pieces t1 to t5 updated as follows are used as the respective column components of the first column to the fifth column. An approximate matrix Dk of columns is generated.

  That is, the updated content vector of the music t1 is (3.3622, 2.9437, 0.7306, 0.4177, 0.9981, 2.8258). The updated content vector of the music piece t2 is (1.0252, 0.7929, 1.8142, 3.2245, 1.0748, 3.4327). The updated content vector of the music t3 is (1.0908, 0.8379, 2.0166, 3.5988, 1.1854, 3.7918). The updated content vector of the musical piece t4 is (1.0652, 0.9030, 0.6816, 1.0083, 0.5341, 1.6224). The updated content vector of the musical piece t5 is (3.6087, 3.1206, 1.3746, 1.5976, 1.3572, 3.9869).

  When the approximate matrix Dk of FIG. 16 is supplied from the LSA computing unit 20 to the content recommendation unit 23, the process proceeds to step S64.

  In step S64, the content recommendation unit 23 determines the evaluation value of each content. In step S65, the content recommendation unit 23 recommends content based on the determination result. Thus, the “content re-evaluation process by LSA” is completed.

  Note that the determination method of the evaluation value of the content in step S64 is not particularly limited, and various determination methods can be employed. For example, for each content vector, if the “evaluation” component in the approximate matrix Dk satisfies the following first to third conditions, it is determined that the corresponding content should be recommended to the user. A determination method can be adopted. Furthermore, based on this determination method, it is possible to adopt a determination method that considers the degree of change in user's temporal preference and weights some old content to avoid recommending the most recently experienced content. is there.

  The first condition refers to a condition that the value of the “evaluation” component in the approximate matrix Dk is larger than the corresponding component value of the original metadata matrix D.

  The second condition refers to a condition that the value of the “evaluation” component in the approximate matrix Dk is greater than a predetermined threshold value.

  The third condition is the correspondence between the above-described feature difference calculated from the value of the “evaluation” component in the approximate matrix Dk or the value of the evaluation value component in the approximate matrix Dk and the original metadata matrix D This indicates a condition that the above-described feature difference calculated from the difference value or the division value with respect to the component value is larger than a predetermined threshold value.

  Specifically, for example, in this case, it is assumed that the second condition is adopted and 2.5 is set as the threshold value. In this case, the contents (songs) whose “evaluation” component value in the approximate matrix Dk is larger than 2.5 are the music t1, the music t2, the music t3, and the music t5. Therefore, in step S64, it is determined that the music t1, the music t2, the music t3, and the music t5 are contents to be recommended, and in step S65, the music t1, the music t2, the music t3, and the music t5 are recommended. .

  The points to be noted here are as follows.

  That is, for example, when attention is paid to the music piece t1, the evaluation value of the music piece t1, that is, the value of the “evaluation” component is a low value of 2, as shown in FIG. Also, since the music t1 is not particularly similar to the music t2 to t4, as shown in FIG. 14, the value of the “evaluation” component of the music t1 after being updated by the LSA calculation is 1.8780, The value is smaller than the threshold 2.5. Therefore, the music t1 was not recommended before the user viewed the new music t5.

  However, after that, when the user views the new music t5, the new music t5 is highly evaluated. That is, as shown in FIG. 15, the component value of the “evaluation” of the music t5 is a high value of 5. The music t5 is most similar to the music t1 among the music t1 to the music t4. Therefore, when the LSA calculation is performed on the metadata matrix of FIG. 15 including the music t5, the high evaluation value of the music t5 is similar to the music t5 based on the relevance of the metadata (music feature). The component value of “evaluation” of the music t1 being played is also updated to a high value of 2.83. Therefore, as described above, the music t1 that was not recommended because it was once given a low evaluation (the possibility that the recommendation was forgotten is high) is based on the recent user interest, that is, the music t5. It is possible to recommend to the user again based on the high evaluation of the user.

  As described above, in the fourth embodiment, the approximate matrix Dk is updated so as to include the content vector of the new content, whereby the existing content is reevaluated (evaluation value is updated). As a result, it is possible to make a recommendation according to the current user's preference from content that has not been recommended in the past, that is, content that has not been recommended and has been discarded in the past. . In other words, it is possible to cope with a change in preference over time.

  (Fifth embodiment)

  Next, a fifth embodiment will be described.

  As described above, the content vector of content is a vector based on metadata. When a large number of metadata is used as the basis of the content vector, naturally, metadata having different properties are often mixed. For example, in many cases, there are metadata with different degrees of influence on others or different degrees of influence from others, such as metadata that is not affected by others due to its nature.

  However, the conventional content recommendation does not take into account the difference in the properties of these metadata, for example, the degree of influence on others and the degree of influence from others, and as a result, it is not always necessary to select content appropriate for the user. There was a problem of not being able to recommend.

  For example, the various algorithms used for metadata weighting are not suitable for metadata of all properties, but are suitable for metadata of one property, but not suitable for metadata of another property. Most cases. However, conventionally, all metadata is weighted using the same algorithm regardless of the difference in properties, and content recommendation is performed using such weighted metadata. In such a case, there is a problem that the content is not always suitable for the user.

  In order to solve this problem, the present inventor has invented the above-described fifth processing, that is, “recommendation processing by hybrid of LSA and other methods”.

  The outline of the fifth process will be described below.

  As described above, the metadata can be classified into several types according to the characteristics, and an appropriate weighting algorithm may be different for each type.

  In such a case, the information processing system or information processing apparatus of the fifth embodiment (hereinafter simply referred to as “device” in the description of the outline of the fifth processing) performs the matrix weighting process used for matching with the metadata. Execute separately for each type.

  The apparatus performs content matching processing using the matrix thus weighted. This makes it possible to perform a more appropriate matching process as compared to the conventional case.

  The apparatus can also change the weight by multiplying the component value calculated by the corresponding algorithm by a predetermined coefficient for every two or more algorithms.

  For example, it is assumed that the content is an e-mail, and the words in the e-mail, the transmission / reception time zone, the exchange partner, and the location are adopted as the metadata. In this case, for example, the device classifies the words in the e-mail among the metadata into the first type, and sets the other three elements, that is, the transmission / reception time zone, the exchange partner, and the location to the second type. Classify the type.

  Next, the apparatus generates a metadata matrix, which is divided into a first submatrix composed of components corresponding to the first type of metadata and a first submatrix composed of components corresponding to the second type of metadata. Partition into 2 sub-matrices.

  Next, for example, the apparatus executes a weighting process for weighting with a general weighting algorithm such as TF / IDF for the first submatrix, while for example the LSA for the second submatrix. The weighting process is executed by the second weighting algorithm. Note that the combination of algorithms at this time is not limited to this example, and it goes without saying that any combination may be used.

  Then, the apparatus combines the first partial matrix and the second partial matrix weighted with different algorithms as described above, and performs a matching process using a matrix obtained as a result (hereinafter referred to as an approximate composite matrix). Do.

  The outline of the fifth process, that is, the “recommendation process using a hybrid of LSA and another method” has been described above.

  Hereinafter, metadata that is classified into the second type is referred to as a context. That is, in this specification, the context refers to all of the internal state and external state of the user. The user's internal state refers to the user's physical condition, emotion (feeling or psychological state), or the like. Further, the external state of the user is distributed in the spatial direction or the temporal direction around the user in addition to the spatial or temporal arrangement position of the user (the temporal arrangement position indicates, for example, the current time). It also refers to a predetermined state (or distributed in any direction).

  Next, referring to FIG. 17 and FIG. 18, the information processing system or information processing apparatus of the fifth embodiment, that is, the information processing system or information processing apparatus that executes “recommendation processing by hybrid of LSA and other methods” Will be described.

  FIG. 17 illustrates a functional configuration example of the information processing system or the information processing apparatus according to the fifth embodiment.

  In other words, blocks necessary for executing the “recommendation process by hybrid of LSA and multi-method” are extracted from all the blocks of the user interface unit 11 to the content recommendation unit 23 in FIG. FIG. 17 is a diagram arranged according to the flow of information during the execution of the “recommendation process using a hybrid method and multi-method”. Therefore, the description of each block shown in FIG. 17 has been described above with reference to FIG.

  Although omitted in the example of FIG. 17, the information transmission unit 24 of FIG. 1 is actually arranged in each arrow connecting the two blocks, that is, between the two blocks. become.

  FIG. 18 is a flowchart for explaining an example of “recommendation processing by hybrid of LSA and other methods”. Therefore, an example of “recommendation process by hybrid of LSA and multi-method” will be described below with reference to the flowchart of FIG.

  Here, for example, a first type of metadata group M1 and a second type of metadata group M2 that is different from the first type are adopted, and among the metadata group M1 and the metadata group M2 It is assumed that there is no influence in the opposite direction, although one is affected from the other.

  Specifically, for example, when a song is processed as content, for example, the feature amount of the song can be adopted as the metadata group M1, and the place, time, and situation where the user experiences the content A context such as emotion can be adopted as the metadata group M1. This is because features and contexts are naturally different, and even if the context affects the impression (features) of music, the music (features) does not directly affect the context. is there.

  Here, the direction of influence is set to the direction from the metadata group M2 to the metadata group M1, for example, according to the example of the content and the feature amount. In this case, of all the metadata, there are s types classified into the metadata group M1, and t types classified into the metadata group M2.

  Further, it is assumed that there are n contents to be processed. That is, it is assumed that s + k pieces of metadata are assigned to each of the n pieces of content.

  In this case, as a result of the processing of the matrix generation unit 18 in step S81 of FIG. 18, the matrix A represented by the following equation (7) is generated as the metadata matrix D.

・・・（７）

... (7)

  In Expression (7), m1u, v (u = 0 to s-1, v = 0 to n-1) is metadata added to the v-th content, and is classified into the metadata group M1. Of the s types of metadata, component values corresponding to the u-th metadata are shown. M2w, x (w = 0 to t−1, x = 0 to n−1) is metadata added to the x-th content, and t types of metadata classified into the metadata group M2. Of the data, the component value corresponding to the w-th metadata is shown.

  In step S82, the matrix generation unit 18 partitions the metadata matrix into two partial matrices. That is, in this case, in step S82, the matrix generation unit 18 divides the matrix into the partial matrix Mt1 and the partial matrix Mt2 as shown in the rightmost side of Expression (7).

  The partial matrix Mt1 is a matrix composed of matrix components for s rows from the top of the matrix A, that is, m1u, v (u = 0 to s-1, v = 0 to n-1) as component values. As a matrix. Therefore, the partial matrix Mt1 is a matrix of s rows and n columns.

  On the other hand, the partial matrix Mt2 is a matrix composed of matrix components for t rows from the bottom of the matrix A, that is, m2w, x (w = 0 to t−1, x = 0 to n− 1) as a component value. Therefore, the partial matrix Mt1 is a matrix with t rows and n columns.

  In step S83, the weighting processing unit 19 performs weighting on each of the two sub-matrices.

  In step S84, the LSA calculation unit 20 performs an LSA calculation on at least one of the two partial matrices.

  Note that performing LSA operation on a submatrix here includes naturally performing an LSA operation on a single submatrix to generate an approximate matrix of a single submatrix, and the entire metadata matrix. The LSA operation is performed on the result, and only the component corresponding to the target submatrix is used in the approximated matrix of the metadata matrix obtained as a result.

  The latter will be specifically described. For example, in this case, when the LSA operation is performed on the entire metadata matrix A of Expression (7), a matrix A ′ represented by the following Expression (8) is generated as an approximate matrix of the metadata matrix A. It will be.

・・・（８）

... (8)

  In this case, the matrix generation unit 18 partitions the approximate matrix A ′ in exactly the same manner as the processing in step S82, that is, the approximate matrix A ′ is similar to the partition of the metadata matrix A into two partial matrices Mt1 and Mt2. Is also divided, two sub-matrices Mt1 ′ and Mt2 ′ are obtained as shown in Expression (8).

  The partial matrix Mt1 ′ is a matrix composed of matrix components for s rows from the top of the approximate matrix A ′, that is, m1u, v (u = 0 to s−1, values updated by LSA calculation). It is a matrix having v = 0 to n-1) as component values. Accordingly, the partial matrix Mt1 'is also a matrix of s rows and n columns.

  On the other hand, the submatrix Mt2 ′ is a matrix composed of matrix components for t rows from the bottom of the approximate matrix A ′, that is, m2w, x (w = 0) whose value is updated by the LSA operation. To t-1, x = 0 to n-1) as component values. Therefore, the partial matrix Mt1 'is also a matrix of t rows and n columns.

  In this case, for example, if the partial matrix Mt1 is targeted, the process of step S84 results in the partial matrix Mt1 'of Expression (8).

  In step S85, the matrix generation unit 18 generates an approximate composite matrix by combining the two partial matrices.

  For example, in this case, a matrix B represented by the following equation (9) is generated as an approximate synthesis matrix.

・・・（９）

... (9)

  In Equation (9), the submatrix Mt1 'is the same matrix as that of Equation (8) described above. The partial matrix Mt2 is a matrix weighted by the process of step S83 with respect to that of the above-described equation (7).

  When the approximate composition matrix B is supplied to the content recommendation unit 23 and the content recommendation unit 23 requests the vector calculation unit 22 to perform a matching process, the process proceeds to step S86.

  In step S86, the vector calculation unit 22 performs content matching processing using the approximate synthesis matrix B. Specifically, for example, in step S86, the vector calculation unit 22 generates a UPV from each column component of the approximate synthesis matrix, that is, from the content vector that the user gave a high evaluation of among the content vectors. The vector calculation unit 22 calculates the similarity between the UPV and the existing content vector, selects the content with the highest similarity (or an arbitrary number of content from the highest), and sends the selection result to the content recommendation unit 23. Notice.

  Then, in step S87, the content recommendation unit 23 recommends the content notified from the vector calculation unit 22. That is, in step S87, the content recommendation unit 23 acquires the content to be recommended from the content recording unit 15, and presents it to the user via the user interface unit 11.

  Thus, the “recommendation process by hybrid of LSA and other methods” is completed.

  In the following, “recommendation processing by hybrid of LSA and other methods” will be further described.

  As described above, the approximate matrix of the metadata matrix A in Expression (7) is the matrix A ′ in Expression (8). The two partial matrices Mt1 'and Mt2' partitioned from the approximated matrix A 'influence each other by dimensional compression on the metadata matrix A in Expression (7).

That is, for example, in the content corresponding to column c of the metadata matrix A, the i-th metadata weight (component value) m1 _{i, c} in the metadata group M1 and the j-th metadata in the metadata group M2 Assume that both of the weights (component values) m2 _{j, c} are large. That is, assume that these two metadata co-occur. In this case, if the weight (component value) of the i-th metadata in the metadata group M1 is large and the weight (component value) of the j-th metadata in the metadata group M2 is small in other content, the LSA The weight (component value) of the jth metadata is raised due to the nature of dimension compression by singular value decomposition of the operation. The same applies when the relationship between the metadata group M1 and the metadata group M2 is reversed.

  For example, when the document is content and the word is metadata, the mutual influence between the metadata group M1 and the metadata group M2 is as described in the first embodiment and the second embodiment. It is shown to be useful as a weighting considering the co-occurrence relationship of words.

  However, in the fifth embodiment, it is assumed that only the influence from the metadata group M2 to the metadata group M1 exists and there is no influence in the opposite direction. Therefore, when such a premise is made, there is a desire to use only the influence from the metadata group M2 to the metadata M1 for weighting.

  Therefore, in order to realize this requirement, in the fifth embodiment, the approximate synthesis matrix B represented by the above-described equation (9) is used as the weighted metadata matrix.

  In the approximate composite matrix B of Equation (9), the lower partial matrix Mt2 is the same as that of step S83 with respect to the metadata matrix A before dimension compression, that is, the metadata matrix A of Equation (7), as described above. It is a submatrix at the top of a matrix weighted by processing. Further, in the approximate synthesis matrix B of Expression (9), the upper partial matrix Mt1 'is the upper partial matrix of the approximate matrix B of Expression (8).

  That is, in the approximate synthesis matrix B of Equation (9), the upper partial matrix Mt1 ′ is a weighting matrix in consideration of the influence of the metadata group M1 from the metadata group M2, while the lower partial matrix Mt2 Is a weighting matrix that is not affected by the metadata group M1.

  Accordingly, it can be said that the approximate synthesis matrix B is a weighted metadata approximation matrix in which only the one-way influence from the metadata group M2 to the metadata group M1 is considered.

  Here, the matrix generation unit 18 further performs weighting such as tf · idf on the partial matrix Mt2 below the approximate synthesis matrix B, or generates a detailed partial matrix further dividing the partial matrix Mt2. Different weights can be applied to each of the sub-matrices. Note that the weighting in this case includes recursive application of singular value decomposition that realizes only one influence as described above.

  By the way, in the above-described example, it is assumed that the metadata group M1 and the metadata group M2 affect only one direction from one to the other, but the metadata group M1 and the metadata group M2 are There are many cases where you want to consider co-occurrence relationships completely independently.

  In such a case, in the process of step S84, the LSA computing unit 20 performs singular value decomposition individually on each of the partial matrix Mt1 and the partial matrix Mt2 of the equation (7) weighted in the process of step S83. Can be applied. This is the meaning of “at least” in “at least one of the two sub-matrices” in step S84.

  That is, in the process of step S84, the LSA operation unit 20 individually executes each of the singular value decompositions shown in the following equations (10) and (11).

・・・（１０）

... (10)

・・・（１１）

(11)

Then, as shown in the following formulas (12) and (13), the LSA calculation unit 20 compresses the partial matrix Mt1 and the partial matrix Mt2 into the approximate partial matrix Mt1 _k1 compressed in the k1 dimension and the k2 dimension, respectively. Each of '' and approximate submatrix Mt2 _k2 '' can be generated.

・・・（１２）

(12)

・・・（１３）

(13)

  Therefore, the matrix generation unit 18 can generate the approximate synthesis matrix A ″ represented by the following equation (14) by the process of step S85.

・・・（１４）

(14)

  Thus, the approximate synthesis matrix A ″ is a weighted metadata approximation matrix in which the metadata group M1 and the metadata group M2 do not affect each other, but the co-occurrence relationship and the like are taken into account in each of them.

  As described above, with reference to FIGS. 17 and 18, the information processing system or information processing apparatus according to the fifth embodiment, that is, the information processing system or information processing apparatus that executes “recommendation processing by hybrid of LSA and other methods”. explained.

  In the fifth embodiment, the metadata group M1 and the metadata group M2 are weighted in consideration of the interrelationships inside each other, or only the influence from the metadata group M2 on the metadata group M1 or the metadata group It is possible to perform weighting based on only the influence from M1 to the metadata group M2. As a result, more appropriate matching processing can be executed as compared with the conventional case, and accordingly, more appropriate content recommendation can be performed as compared with the conventional case.

  The first to fifth embodiments have been described above.

  By the way, the series of processes described in the first to fifth embodiments can be executed by hardware, but can also be executed by software.

  In this case, the information processing apparatus of FIG. 1 can be configured by, for example, a personal computer as shown in FIG.

  In FIG. 19, a CPU (Central Processing Unit) 101 executes various processes according to a program recorded in a ROM (Read Only Memory) 102 or a program loaded from a storage unit 108 to a RAM (Random Access Memory) 103. To do. The RAM 103 also appropriately stores data necessary for the CPU 101 to execute various processes.

  The CPU 101, ROM 102, and RAM 103 are connected to each other via a bus 104. An input / output interface 105 is also connected to the bus 104.

  The input / output interface 105 includes an input unit 106 including a keyboard and a mouse, an output unit 107 including a display, a storage unit 108 including a hard disk, and a communication unit 109 including a modem and a terminal adapter. It is connected. The communication unit 109 performs communication processing with other information processing apparatuses via a network including the Internet.

  A drive 110 is also connected to the input / output interface 105 as necessary, and a removable recording medium 111 made of a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is appropriately installed, and a computer program read from them is read. Are installed in the storage unit 108 as necessary.

  When a series of processing is executed by software, a program constituting the software executes various functions by installing a computer incorporated in dedicated hardware or various programs. For example, a general-purpose personal computer is installed from a network or a recording medium.

  As shown in FIG. 19, the recording medium including such a program is distributed to provide a program to the user separately from the main body of the apparatus, and includes a magnetic disk (including a floppy disk) on which the program is recorded. , Removable recording media (packages) consisting of optical disks (including compact disk-read only memory (CD-ROM), DVD (digital versatile disk)), magneto-optical disks (including MD (mini-disk)), or semiconductor memory (Media) 111, but also a ROM 102 on which a program is recorded and a hard disk included in the storage unit 108 provided to the user in a state of being incorporated in the apparatus main body in advance.

  In the present specification, the step of describing the program recorded on the recording medium is not limited to the processing performed in time series along the order, but is not necessarily performed in time series, either in parallel or individually. The process to be executed is also included.

  Further, in the present specification, the system represents the entire apparatus including a plurality of apparatuses and processing units.

It is a functional block diagram which shows the functional structural example of the information processing system to which this invention is applied. FIG. 2 is a functional block diagram showing the information processing system of FIG. 1 from the viewpoint of information flow when executing “unnecessary metadata extraction processing considering co-occurrence relation”. 3 is a flowchart for explaining “unnecessary metadata extraction processing considering co-occurrence relations” executed by the information processing system of FIG. 2. It is an example which shows the process result of the "unnecessary metadata extraction process which considered the co-occurrence relation" of FIG. It is an example which shows the process result of the "unnecessary metadata extraction process which considered the co-occurrence relation" of FIG. It is an example which shows the process result of the "unnecessary metadata extraction process which considered the co-occurrence relation" of FIG. FIG. 2 is a functional block diagram illustrating the information processing system of FIG. 1 from the viewpoint of information flow when executing a “recommendation process considering co-occurrence relationships”. It is a flowchart explaining the "recommendation extraction process in consideration of the co-occurrence relation" performed by the information processing system of FIG. FIG. 2 is a functional block diagram illustrating the information processing system of FIG. 1 from the viewpoint of information flow when executing “recommendation processing using a difference between clustered UPV groups”. 10 is a flowchart illustrating “recommendation processing using a difference between clustered UPV groups” executed by the information processing system of FIG. 9. FIG. 2 is a functional block diagram illustrating the information processing system of FIG. 1 from the viewpoint of information flow when executing “content re-evaluation processing by LSA”. 12 is a flowchart for explaining “content reevaluation processing by LSA” executed by the information processing system of FIG. 11; 12 is an example showing a processing result of “content re-evaluation processing by LSA” in FIG. 12 is an example showing a processing result of “content re-evaluation processing by LSA” in FIG. 12 is an example showing a processing result of “content re-evaluation processing by LSA” in FIG. 12 is an example showing a processing result of “content re-evaluation processing by LSA” in FIG. FIG. 2 is a functional block diagram showing the information processing system of FIG. 1 from the viewpoint of information flow when executing “recommendation processing by hybrid of LSA and other methods”. 18 is a flowchart for explaining “recommendation processing by hybrid of LSA and other method” executed by the information processing system of FIG. It is a block diagram which shows the structural example of the hardware of the information processing apparatus (at least one part of the information processing system of FIG. 1) to which this invention is applied.

Explanation of symbols

  11 user interface unit, 12 user profile storage unit, 13 user dictionary storage unit, 14 general dictionary storage unit, 15 content storage unit, 16 metadata storage unit, 17 metadata acquisition unit, 18 matrix generation unit, 19 weighting processing unit, 20 LSA operation unit, 21 metadata extraction unit, 22 vector operation unit, 101 CPU, 102 ROM, 103 RAM, 108 storage unit, 111 removable recording medium

Claims (4)

Based on N (N is an integer value of 1 or more) metadata associated with at least one of the plurality of contents, M (M is an integer value of 1 or more) of the plurality of contents. Matrix generating means for vectorizing each of the contents and generating a matrix having M vectors obtained as a result thereof as column components or row components as a metadata matrix;
Approximate matrix generation means for generating an approximate matrix of the metadata matrix by performing singular value decomposition on the metadata matrix generated by the matrix generation means;
Each of the M column components or row components indicating each of the M contents of the approximate matrix generated by the approximate matrix generation means is divided as a vector, and each of the vectors divided into M pieces Classifying means for classifying S into a predetermined one of S clusters (S is an integer less than or equal to M);
For each of the S clusters, a representative vector representing the corresponding cluster based on the vector classified into the corresponding cluster by the classification means, and the representative based on the N metadata Representative vector generation means for generating a vector;
One or more sets of representative vectors of two clusters are selected from the representative vectors of the S clusters generated by the representative vector generation means, and two clusters are selected for each of the selected one or more sets. Difference vector generation means for generating a difference vector of the representative vector;
Another content different from the M content is vectorized based on the N metadata, and the resulting vector and at least one of the one or more difference vectors generated by the difference vector generation means An information processing apparatus comprising: similarity calculation means for calculating similarity with one. Recommendation means for determining whether or not the other content is content to be recommended to a user based on at least one of the one or more similarities calculated by the similarity calculation means;
The presentation unit according to claim 1, further comprising: a presenting unit that presents the other content to the user when the other unit determines that the other content should be recommended to the user. Information processing device. Based on N (N is an integer value of 1 or more) metadata associated with at least one of the plurality of contents, M (M is an integer value of 1 or more) of the plurality of contents. A matrix generation step of generating each of the contents as a metadata matrix by vectorizing each of the contents, and generating a matrix having the M vectors obtained as a result thereof as column components or row components;
An approximate matrix generation step of generating an approximate matrix of the metadata matrix by performing singular value decomposition on the metadata matrix generated by the processing of the matrix generation step;
Each of the column component or the row component indicating each of the M contents of the approximate matrix generated by the processing of the approximate matrix generation step is partitioned as a vector, and each of the M partitioned vectors is determined. , A classification step for classifying into a predetermined one of S clusters (S is an integer value less than or equal to M);
For each of the S clusters, a representative vector representing the corresponding cluster based on the vector classified into the corresponding cluster by the processing of the classification step, and based on the N metadata A representative vector generation step of generating the representative vector;
One or more sets of representative vectors of two clusters are selected from the representative vectors of the S clusters generated by the processing of the representative vector generation step, and each of the selected one or more sets includes two A difference vector generation step for generating a difference vector of the representative vector of the cluster;
Other contents different from the M contents are vectorized based on the N pieces of metadata, and the resulting vector and one or more of the difference vectors generated by the processing of the difference vector generation step A similarity calculation step of calculating a similarity with at least one of the information processing method. A program to be executed by a computer,
Based on N (N is an integer value of 1 or more) metadata associated with at least one of the plurality of contents, M (M is an integer value of 1 or more) of the plurality of contents. A matrix generation step of generating each of the contents as a metadata matrix by vectorizing each of the contents, and generating a matrix having the M vectors obtained as a result thereof as column components or row components;
An approximate matrix generation step of generating an approximate matrix of the metadata matrix by performing singular value decomposition on the metadata matrix generated by the processing of the matrix generation step;
Each of the column component or the row component indicating each of the M contents of the approximate matrix generated by the processing of the approximate matrix generation step is partitioned as a vector, and each of the M partitioned vectors is determined. , A classification step for classifying into a predetermined one of S clusters (S is an integer value less than or equal to M);
For each of the S clusters, a representative vector representing the corresponding cluster based on the vector classified into the corresponding cluster by the processing of the classification step, and based on the N metadata A representative vector generation step of generating the representative vector;
One or more sets of representative vectors of two clusters are selected from the representative vectors of the S clusters generated by the processing of the representative vector generation step, and each of the selected one or more sets includes two A difference vector generation step for generating a difference vector of the representative vector of the cluster;
Other contents different from the M contents are vectorized based on the N pieces of metadata, and the resulting vector and one or more of the difference vectors generated by the processing of the difference vector generation step A similarity calculation step for calculating a similarity with at least one of the program.

Images