JP6040137B2 - Item recommendation device, item recommendation method, and item recommendation program - Google Patents

Description

  The present invention is an apparatus for recommending items related to certain information, in particular, an item recommendation apparatus for estimating and recommending items related to programs such as TV and VOD (video on demand) based on the trend of the world, Methods and programs.

  With the digitization of TV and the spread of the Internet, it has become possible not only to watch programs but also to use various services via TV. Specifically, services such as browsing of program related information and shopping are provided (see, for example, Patent Documents 1, 2, and 3).

  In Patent Document 1, it is premised that a program and an item are associated in advance with respect to an item to be recommended. However, it is a costly operation to manually select a related item from a large number of items.

  On the other hand, according to Patent Document 2, related information can be provided at a predetermined timing of video data based on metadata of an electronic program guide (EPG) or closed caption (CC).

  According to Patent Document 3, a method of recommending a program that matches a user's preference by analyzing program information that has been viewed in the past based on the user's program viewing history and grasping the user's preference. Has been proposed.

  Note that the topic direc- tion allocation (LDA) used in topic extraction in the present invention is described in Non-Patent Document 1, and the program topic relevance calculation unit of the present invention is based on the adjacency matrix. Further, Non-Patent Document 2 describes a power iteration technique used when obtaining each feature amount of a topic.

JP 2002-330422 A JP 2013-164770 A JP 2008-199406 A

D. M.M. Blei, A .; Y. Ng, and M.M. I. Jordan. Latent Dirichlet Allocation, in Journal of Machine Learning Research (JMLR), vol. 3, pp. 993-1022 (2003). Lin, F.M. and Cohen, W.M. W. : Power iteration clustering, In Proc. Of ICML, pp. 655-662, 2010.

  However, conventional techniques using program information and viewing history can recommend items based on user preferences, but cannot make recommendations based on trends in the world. For example, you can recommend items such as sports videos and equipment to users who are watching sports programs, but topics that are currently discussed, such as what sports videos and equipment are currently being discussed There was a problem that items related to could not be recommended.

  In addition, many of the conventional techniques including Patent Document 2 search for recommended items based on program information, but there are many cases where the amount of information of programs to be used is insufficient, and many items related to the programs. There is a problem that cannot be recommended. General program information includes EPG distributed with broadcast waves and closed caption CC.

  However, since the EPG is an outline of a program, there is a lack of specific information, and there are many programs that do not include a CC. For example, if you want to recommend shoes and wear to viewers of Hakone Ekiden, the program information for Hakone Ekiden will not include the word shoes, but conversely the shoe metadata will include the word Hakone Ekiden. Therefore, it cannot be recommended by the conventional technology. On the other hand, there is a service for manually increasing program information, but this has a problem that it is very expensive.

  SUMMARY OF THE INVENTION The present invention solves the above-mentioned problems, and an object thereof is to provide an item recommendation device, method, and program capable of recommending an item related to a topic based on a trend in the world and related to a program. There is to do.

  An item recommendation device according to the present invention for solving the above-described problem is an item recommendation device for recommending an item related to a program, and extracts a topic including a feature word from a text stream storage unit in which a set of texts is stored. Topic extraction means, trend score calculation means for calculating the trend score of the topic extracted by the topic extraction means, and storing the trend score together with the topic; and related to program information from the Web information database storing the Web information Program-related Web information extracting means for extracting Web information to be stored and storing the extracted program-related Web information; program information; program-related Web information extracted by the program-related Web information extracting means; and the trend score calculating means Programs and topics based on memorized topics Program / topic relevance calculating means for calculating the relevance of items, item information database storing items and information related thereto, items in the item information database, programs calculated by the program / topic relevance calculating means Program / item relevance calculating means for calculating the relevance of a program and an item based on the topic relevance and the topic trend score calculated and stored by the trend score calculating means, and the program / item relevance calculation Output means for outputting an item recommended in the program based on the degree of association between the program and the item calculated by the means.

  The text stream is assumed to be text with high real-time properties such as a microblog (for example, Twitter (registered trademark)).

  ADVANTAGE OF THE INVENTION According to this invention, the item relevant to the topic of a topic based on the trend of the world and relevant to a program can be recommended. Further, even if the program information has a small amount of information only of EPG and CC, since the program related Web information is extracted and used, it is possible to widely associate programs with items and recommend related items.

The block diagram which shows one embodiment of this invention. The flowchart until it calculates a trend topic from a text stream in the example of one embodiment of the present invention. Explanatory drawing which shows the example of the text stream of the text stream memory | storage part in one example of embodiment of this invention. Explanatory drawing which shows the example of the extracted topic and the calculated trend score in one example embodiment of this invention. The flowchart of the item recommendation based on the trend topic relevant to the program in the example embodiment of the present invention. Explanatory drawing which shows an example of the adjacency matrix prepared when calculating the relevance degree of program relevant information and a topic in one example of this invention. Explanatory drawing which shows an example of item information DB (database) in one embodiment of this invention. Explanatory drawing which shows an example of the recommendation item relevant to the program which the output part in one embodiment of this invention outputs.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings, but the present invention is not limited to the following embodiments. FIG. 1 shows the configuration of an item recommendation device according to an embodiment of the present invention.

  The item recommendation device shown in FIG. 1 includes a text stream storage unit 10, a topic extraction unit 20 as a topic extraction unit, a trend score calculation unit 30 and a topic / trend score storage unit 40 constituting a trend score calculation unit, a Web information DB 50, Program-related Web information extracting unit 60 and program-related Web information storage unit 70 constituting program-related Web information extracting unit, program / topic related level calculating unit 80 as program / topic related level calculating unit, item information DB 90, program / item A program / item relevance calculation unit 100 as relevance calculation means and an output unit 110 as output means are provided.

  The topic extraction unit 20 extracts a topic including a feature word from the text stream storage unit 10 in which a set of text such as a microblog (for example, Twitter) is stored.

  The trend score calculation unit 30 calculates a trend score representing the trend degree of the topic extracted by the topic extraction unit 20, and stores it in the topic / trend score storage unit 40 together with the topic.

  The program-related Web information extraction unit 60 extracts Web information related to program information (EPG (electronic program guide), CC (closed caption) data) from the Web information DB 50 in which Web information is stored, and the extracted program The related Web information is stored in the program related Web information storage unit 70.

  The program / topic relevance calculation unit 80 associates a program with a topic based on program information, program-related Web information stored in the program-related Web information storage unit 70, and topics stored in the topic / trend score storage unit 40. Calculate the degree.

  The item information DB 90 stores items and information related thereto.

  The program / item relevance calculation unit 100 includes the item in the item information DB 90, the relevance between the program and the topic calculated by the program / topic relevance calculation unit 80, and the topic trend stored in the topic / trend score storage unit 40. Based on the score, the degree of association between the program and the item is calculated.

  The output unit 110 outputs an item recommended in the program based on the degree of association between the program and the item calculated by the program / item relevance calculation unit 100.

  The item recommendation device in FIG. 1 is configured by a computer, for example, and includes hardware resources of a normal computer, such as a ROM, a RAM, a CPU, an input device, an output device, a communication interface, a hard disk, a recording medium, and a driving device thereof. .

  As a result of the cooperation between the hardware resource and the software resource (OS, application, etc.), as shown in FIG. 1, the item recommendation device has a text stream storage unit 10, a topic extraction unit 20, a trend score calculation unit 30, and a topic. Trend score storage unit 40, Web information DB 50, program related web information extraction unit 60, program related web information storage unit 70, program / topic relevance calculation unit 80, item information DB 90, program / item relevance calculation unit 100, output Part 110 is implemented.

  The text stream storage unit 10, the topic / trend score storage unit 40, the Web information DB 50, the program-related Web information storage unit 70, and the item information DB 90 are constructed in storage means / storage means such as a hard disk or RAM. .

  Next, the operation of the apparatus configured as described above will be specifically described. The processing in the present embodiment is divided into two types: calculation of a trend topic from a text stream, and item recommendation based on a trend topic associated with a program.

  First, calculation of trend topics from a text stream will be described. FIG. 2 shows a flowchart of processing for calculating a trend topic from a text stream according to an embodiment of the present invention.

  <Step S210> The computer stores the generation time, text information, and text ID in the text stream storage unit 10 for the text stream for a predetermined time. An example of storage is shown in FIG. The stored text stream is deleted after a predetermined time, and a new text stream for a predetermined time is stored again. Here, the text stream is assumed to be text with high real-time properties such as a microblog (for example, Twitter).

  <Step S220> The topic extraction unit 20 extracts the topic zεZ from the text dεD stored in the text stream storage unit 10. Here, D is a text set, Z is a topic set, and topic z is represented by a set of feature words wεW (W is a set of the entire vocabulary). For topic extraction, the topic extraction unit 20 uses a morphological analyzer to divide the text into word sets such as nouns, verbs, and adjectives. For the topic extraction, a well-known topic model such as Late Dichlet Allocation (LDA) (Non-Patent Document 1), which is a typical topic model, can be used. FIG. 4 shows an example of the extracted topic. Each topic has a topic ID and a trend score. A method for calculating the trend score will be described later.

<Step S230> The trend score calculation unit 30 calculates the trend score T _zi of each extracted topic z _i . The trend score is calculated by regarding that the topic has a higher degree of trend as the more text related to the topic is generated. For example, it calculates like the following formula | equation (1).

Here, i and j are topic and text IDs, respectively, and P (z _i | d _j ) is the occurrence probability of topic z _i in text d _j . FIG. 4 shows an example of the trend score calculated for each topic.

  <Step S240> The extracted and calculated topic and trend score are stored in the topic / trend score storage unit 40. An example of storage is as shown in FIG.

  Next, item recommendation based on trend topics related to programs will be described. FIG. 5 shows a flowchart of a process for recommending an item based on a trend topic associated with a program according to an embodiment of the present invention.

  <Step S310> The Web information DB 50 stores, for example, Wikipedia (registered trademark) articles obtained by crawling the Internet, Tweet (registered trademark) with Twitter program hashtags, and the like. Web information related to the programs EPG and CC is transmitted according to the request from the unit 60. The program-related Web information extraction unit 60 divides the text of the Web information in the EPG, CC, and Web information DB 50 into word units by a morphological analyzer, for example, and then represents the characteristics of the text based on the word information that appears for each text. Build a vector. The relevance of these feature vectors is calculated based on the cosine similarity and the like, and Web information related to the program is extracted.

  <Step S320> Information transmitted from the Web information DB 50 in response to a request from the program-related Web information extraction unit 60 is stored in the program-related Web information storage unit 70. Every time EPG or CC, which is program information being viewed, changes with time, the program-related Web information stored in the program-related Web information storage unit 70 is updated.

  <Step S330> tv indicating that the program / topic relevance calculation unit 80 indicates a program (EPG / CC text), program-related Web information aεA (A is the program-related Web information stored in the program-related Web information storage unit 70) Set), and the degree of association with the topic zεZ stored in the topic trend score storage unit 40 is calculated. In order to widely associate the program tv and the topic z, the program-related Web information a is used as a bridge connecting the information. Specifically, for example, an adjacency matrix C is calculated assuming a graph representing the relationship between the program tv, the program-related Web information a, the topic z, and the feature word wεW. Based on the adjacency matrix C, each feature amount is calculated, and the degree of association is calculated by comparing each feature amount.

The adjacency matrix C expresses the relationship between program information, topics, and feature words as one matrix as shown in FIG. The adjacency matrix C is constructed by the following procedure. A matrix C = (c _j , _k ) of (1+ | W | + | Z | + | A |) × (1+ | W | + | Z | + | A |) is prepared as an adjacency matrix. j represents a program when j = 1, and represents a feature word when 2 ≦ j ≦ | W | +1. Also, a topic is represented when | W | + 2 ≦ j ≦ | W | + | Z | +1, and a program when | W | + | Z | + 2 ≦ j ≦ | W | + | Z | + | A | +1 Represents related Web information. The same applies to k. For example, c _{j, k} = 1 (j = 1, 2 ≦ k ≦ | W | +1) represents that the value representing the relationship between the program in the matrix and the feature word is 1.

The value of each element of the matrix is calculated as follows: For example, the word-topic probability P (w _k | z _j ) calculated by the LDA of Non-Patent Document 1 described above is substituted into the matrix value c _{j, k} representing the combination of the topic z and the feature word w. The matrix value c _{j, k} representing the combination of the program tv and the program-related Web information a and the feature word w is, for example, tf / σ based on the word frequency tf in which the word w occurs in the text of tv and a. Is assigned. Here, σ is a constant. Finally, 0 is assigned to the element c _{j, k} that has not yet been assigned a value.

  In order to calculate the degree of association between the program tv and the topic z, each feature amount is calculated based on the constructed adjacency matrix C. The feature amount is calculated by constructing a graph Laplacian from the adjacency matrix and performing eigenvalue decomposition as in the following equation (2), for example.

Here, B is a diagonal matrix having the sum of each row of the adjacency matrix C as diagonal component values, P is a matrix holding eigenvectors as column components, and ∧ is a matrix holding eigenvalues as diagonal components. When the vector of the row component of P is P _i , for example, the feature quantity of the program tv located at the first component in the adjacency matrix is P ₁ . This feature amount represents the intrinsic structure of the graph based on the adjacency matrix C, and the more similar the node, the more similar the feature amount. Therefore, the degree of association can be calculated by comparing these feature amounts. The eigenvector P can be calculated using a known technique such as Power Iteration described in Non-Patent Document 2.

Based on the feature amount calculated as described above, the degree of association R _{tv, zi} between the program tv and the topic z is calculated by the following equation (3) based on the cosine similarity, for example.

  Here, i represents a topic ID.

  <Step S340> The program / item relevance calculation unit 100 calculates the relevance between the item lεL stored in the item information DB 90 and the program tv in consideration of the relevance with the trend topic z. FIG. 7 shows a storage example of the item information DB 90. For each item, item ID, genre, item name, item metadata, etc. are stored.

First, the degree of association R _{lk, zi} between items and topics is calculated. The relevance R _{lk, zi} can be calculated as shown in the following equation (4) using, for example, a result (topic z _i ) extracted by applying the topic model (LDA) to the above-described text stream.

Here, k indicates an item ID. N _lk (w) represents the number of occurrences of the word w in the item metadata l _k .

Using the topic trend score T _zi , the program-topic relevance level R _{tv, zi} , and the item-topic relevance level R _{lk, zi} , the relationship between the program tv and the item l _k is finally _obtained . The degree R _{tv, lk} is calculated _, for example, by the following equation (5).

  Here, λ is a constant that determines the specific gravity of the trend score and relevance.

<Step S350> The output unit 110 includes a program tv and item l _k of relevance R _tv, based on _lk, and output items in the descending order of the degree of relation. Here, output is a concept including display on a display, printing on a printer, sound output, transmission to an external device, and the like. The output unit 110 may be realized by output device driver software, or output device driver software and an output device. A specific example of output is shown in FIG. FIG. 8 shows an example of recommending a topical shoe to a user who is viewing a marathon program. By identifying trend topics in SNS (Social Networking Service), it is possible to recommend trend items that are relevant by detecting, for example, topical products and associating marathons and shoes with program-related Web information .

  In addition, a part or all of the functions of each unit in the item recommendation device of the present embodiment can be configured by a computer program, and the program can be executed using the computer to realize the present invention. It goes without saying that the procedure in the item recommendation method can be constituted by a computer program, and the program can be executed by the computer, and the computer-readable recording medium, for example, FD, can be realized by the computer. (Floppy (registered trademark) Disk), MO (Magneto-Optical disk), ROM (Read Only Memory), memory card, CD (Compact Disk) -ROM, DVD (Digital Versati) e Disk) -ROM, CD-R, CD-RW, HDD, and recorded in a removable disk, or stored, it is possible or distribute. It is also possible to provide the above program through a network such as the Internet or electronic mail.

  The present invention is not limited to the above-described embodiments, and various modifications and applications can be made within the scope of the claims.

DESCRIPTION OF SYMBOLS 10 ... Text stream memory | storage part 20 ... Topic extraction part 30 ... Trend score calculation part 40 ... Topic / trend score memory | storage part 50 ... Web information DB
60 ... Program-related Web information extraction unit 70 ... Program-related Web information storage unit 80 ... Program / topic relevance calculation unit 90 ... Item information DB
100: Program / item relevance calculation unit 110 ... Output unit

Claims (5)

An item recommendation device for recommending items related to a program,
Topic extraction means for extracting a topic including a feature word from a text stream storage unit in which a set of texts is stored;
A trend score calculating means for calculating a trend score representing a trend degree of a topic extracted by the topic extracting means, and storing the trend score together with the topic;
Program-related Web information extracting means for extracting Web information related to program information from a Web information database in which Web information is stored, and storing the extracted program-related Web information;
Program / topic relevance calculating means for calculating relevance between a program and a topic based on program information, program related Web information extracted by the program related Web information extracting means, and a topic stored by the trend score calculating means; ,
An item information database storing items and information about them,
Based on the item in the item information database, the relevance between the program and the topic calculated by the program / topic relevance calculating means, and the trend score of the topic calculated and stored by the trend score calculating means, Program / item relevance calculation means for calculating relevance;
Output means for outputting an item recommended in the program based on the degree of association between the program and the item calculated by the program / item relevance calculating means;
An item recommendation device comprising: The trend score calculation means sets each text of the text set D stored in the text stream storage unit as d _j , each extracted topic as z _i, and the trend score Tz _i of each topic z _i ,

Calculated by the following formula:
The program / topic relevance calculation means includes tv as the program information, aεA as program-related Web information (A is a set of program-related Web information), zεZ (Z is a topic set), and the features. A word is assumed as w∈W (W is a set of whole vocabulary), and a graph representing the relationship with these is assumed, and (1+ | W | + | Z | + | A |) × (1+ | W | + | Z | + | A |) is constructed, an adjacency matrix C = (c _j , _k ) is constructed, feature quantities of the program tv and topic z are obtained based on the adjacency matrix C, and the feature quantities are compared with each other. Calculate topic relevance R _{tv, zi} ,
The program / item relevance calculation means has a topic-item probability P (z _i | l _k ), a topic-word probability P (z _i | w), a word-item probability P (w | l _k ), The occurrence probability of the topic z _i in the text d _j is P (z _i | d _j ), the number of occurrences of the word w in the item metadata l _k is n _lk (w), and the item-topic relevance R _{lk, zi} The

Calculated by the following formula:
Furthermore, a constant λ that determines the specific gravity of the trend score and the relevance degree, a trend score T _{zi of the} topic calculated by the trend score calculation means, and a relevance degree R of the program and topic calculated by the program / topic relevance calculation means _{Based on tv, zi} and the calculated item-topic relevance R _{lk, zi} , the program-item relevance R _{tv, lk} is

The item recommendation device according to claim 1, wherein the item recommendation device is calculated by the following formula. An item recommendation method for recommending items related to a program,
A topic extraction means for extracting a topic including a feature word from a text stream storage unit in which a set of texts is stored;
A trend score calculating means for calculating a trend score representing a trend degree of the topic extracted by the topic extracting means, and storing the trend score together with the topic; and
A program-related Web information extracting unit, wherein a program-related Web information extracting unit extracts Web information related to program information from a Web information database in which Web information is stored, and stores the extracted program-related Web information;
The program / topic relevance calculating means calculates the relevance between the program and the topic based on the program information, the program related Web information extracted by the program related Web information extracting means, and the topic stored by the trend score calculating means. A program / topic relevance calculation step,
The program / item relevance calculating means includes an item in an item information database storing items and information related thereto, the relevance between the program and the topic calculated by the program / topic relevance calculating means, and the trend score calculating means. A program / item relevance calculating step for calculating a relevance between a program and an item based on a trend score of a topic calculated and stored;
An output step of outputting an item recommended in the program based on the degree of association between the program and the item calculated by the program / item relevance calculating unit;
Item recommendation method characterized by comprising. In the trend score calculation step, each text of the text set D stored in the text stream storage unit is d _j , each extracted topic is z _i, and the trend score Tz _i of each topic z _i is

Calculated by the following formula:
In the program / topic relevance calculation step, the program information is tv, the program-related Web information is aεA (A is a set of program-related Web information), the topic is zεZ (Z is a topic set), and the features A word is assumed as w∈W (W is a set of whole vocabulary), and a graph representing the relationship with these is assumed, and (1+ | W | + | Z | + | A |) × (1+ | W | + | Z | + | A |) is constructed, an adjacency matrix C = (c _j , _k ) is constructed, feature quantities of the program tv and topic z are obtained based on the adjacency matrix C, and the feature quantities are compared with each other. Calculate topic relevance R _{tv, zi} ,
In the program / item relevance calculation step, the topic-item probability is P (z _i | l _k ), the topic-word probability is P (z _i | w), the word-item probability is P (w | l _k ), The occurrence probability of the topic z _i in the text d _j is P (z _i | d _j ), the number of occurrences of the word w in the item metadata l _k is n _lk (w), and the item-topic relevance R _{lk, zi} The

Calculated by the following formula:
Furthermore, a constant λ that determines the specific gravity of the trend score and the relevance degree, a trend score T _{zi of the} topic calculated by the trend score calculation means, and a relevance degree R of the program and topic calculated by the program / topic relevance calculation means _{Based on tv, zi} and the calculated item-topic relevance R _{lk, zi} , the program-item relevance R _{tv, lk} is

The item recommendation method according to claim 3, wherein the item recommendation method is calculated by the following formula.   An item recommendation program for causing a computer to function as each means according to claim 1.

Images