JP5416063B2 - Content recommendation device, content recommendation program and recording medium thereof - Google Patents

Description

  The present invention relates to a technique for recommending another content related to a content, such as a document, an image, music, or video, to a user.

  With the spread of the Internet and the like, the total amount of content that can be viewed by users is increasing. However, the content that the user can browse in a certain time is limited, and it has become quite difficult to find the content that the user is interested in from a large amount of content. One approach to solving this problem is recommendation. Recommendation is a method for facilitating user discovery of content by allowing users to browse more content by presenting other related content when browsing or evaluating a certain content. It is.

  The methods for recommending content can be roughly divided into two types, collaborative filtering and content-based filtering.

  Collaborative filtering is a technique for analyzing whether a user or content is similar by using a user's browsing history or evaluation history for the content and making a recommendation. Collaborative filtering can provide highly accurate recommendations when a large number of user histories are available, but has the disadvantage that the accuracy decreases when the amount of histories is small.

  On the other hand, content-based filtering is a method for recommending similar contents by comparing the contents. Since content-based filtering can be recommended without the history of many users, there is an advantage that content with a small number of viewers can be targeted for recommendation.

  As research on content-based filtering, there is a method of measuring the frequency of appearance of attributes assigned to content highly evaluated by users and recommending content with attributes with high appearance frequency (see Patent Document 1). .

JP 2009-205418 A

  However, in the technique described in Patent Document 1, since the operation history of the user to be recommended is used instead of using the history of a plurality of users, the user to be recommended is recommended in order to make a recommendation with high accuracy. You need to browse a lot of content. Therefore, there is a problem that it is not possible to make an appropriate recommendation for a user with a small number of browsing.

  The present invention solves such a problem, and even when there is no large number of browsing history information, it is possible to recommend another appropriate content related to the content to the user who has browsed the content. The purpose is to be able to.

  The present invention is to solve the above-mentioned problem, and calculates the weight for each keyword assigned to the content by analyzing the relationship between the content set and the keyword assigned to them, and By comparing the weights, content with similar keywords with high weights in each content is recommended.

  That is, the present invention discovers and recommends another related content for a certain content by analyzing each keyword assigned to the content. In keyword analysis, weighting is performed by analyzing the tendency of keywords to be assigned to content. Then, based on the weighted keyword, scoring is performed for each content in the content set that is a candidate for recommendation. The more highly weighted keywords are given, the higher the related score of the content. The content recommendation device recommends content having a high score in scoring.

  The keywords assigned to the content are weighted as follows. The strength of the relationship can be obtained from the co-occurrence relationship between keywords. The weight of each keyword assigned to the content is obtained from the strength of the relationship between the keywords. Alternatively, the weight of the keyword itself is determined from the bias of the keyword co-occurrence appearance. The sum of the weights of keywords common to both content a and content b is used as a related score between the two contents. In this way, content with a high related score is selected and recommended to the user.

  Clustering of a content set having a related score equal to or higher than a threshold value may be performed, and the clustered result may be presented to the user.

  In addition, user browsing information storage means for storing the user's content browsing information is provided, and in the weighting of keywords, the user browsing information storage means is referred to, and by receiving input of a plurality of contents browsed by the user, a plurality of contents can be stored. The keyword weight may be calculated based on the assigned keyword information.

  Further, it is possible to select a keyword that the user is interested in or not interested by input from the user, and to provide keyword weight changing means for changing the weight of the selected keyword so that the keyword weight can be changed.

  According to the present invention, it is possible to calculate the relevance between contents by using only the information about the relation between the content set and the keyword set assigned to them. It is possible to recommend another content related to a keyword.

  In addition, the present invention calculates the strength of relevance between content and keyword for each keyword assigned to the content, and uses it as the weight of the keyword, so that the content strongly related to the content is similar. Can be recommended.

  In addition, the present invention can reduce the amount of calculation required for recommending content by calculating the concreteness indicating whether a keyword can narrow down the topic for each keyword and using it as a keyword weight. it can.

  In addition, according to the present invention, it is possible to recommend a wide variety of contents by clustering several contents with high related scores, collecting them, and presenting them to the user.

  Further, according to the present invention, by recommending content using information of a plurality of contents, it is possible to recommend content reflecting personal preference to a user who has browsed many contents. .

  Further, according to the present invention, by allowing the user to freely change the weight of the keyword, it is possible to recommend other content when the recommended content is not satisfied by the user.

It is a figure which shows the structural example of the content recommendation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the outline | summary of a content recommendation process. It is a figure which shows the example of the information stored in the keyword storage device. It is a figure which shows the example of the information stored in the keyword storage device. It is a figure which shows the example which calculates the strength of the relationship between keywords. It is a figure which shows the example which calculates the frequency | count of cooccurrence between keywords. It is a figure which shows the example of the concrete calculation of a keyword. It is a figure which shows the example of the content recommendation apparatus using a user's browsing history. It is a figure which shows the example of the content recommendation apparatus which can change a recommendation content by user operation.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram schematically showing a content recommendation device according to an embodiment of the present invention.

  The content recommendation device 10 includes a keyword storage device 100, a content information input unit 11, a keyword weighting unit 12, a content related score calculation unit 13, a recommended content determination unit 14, and a content information output unit 15. The keyword storage device 100 stores a content information management table 101, a keyword information management table 102, and a keyword / content relationship management table 103.

  The input / output device 20 is a display, a keyboard, and other peripheral devices used by a user who browses content, but may be a terminal connected via a network. In addition, a system using content recommendation may be used. Hereinafter, a person or system using the content recommendation device 10 is referred to as a “user”.

  The content recommendation device 10 has been developed in, for example, a CPU (Central Processing Unit), storage means such as ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), and storage means. Includes programs. The program describing the operation of the component shown in FIG. 1 can be executed on a computer used as the content recommendation device 10 or can be executed as a service via a network or the like. The keyword storage device 100 may be a program that acquires content and keywords through an API (Application Program Interface) or the like.

  FIG. 2 is a flowchart showing an outline of content recommendation processing by the content recommendation device 10.

  The content recommendation device 10 determines the recommended content based on the keyword information given to the content as follows. First, in step S <b> 1, the content information input unit 11 acquires information on the content accessed by the user, and the keyword weighting unit 12 acquires the keyword set assigned to the content from the keyword storage device 100. Next, in step S2, the keyword weighting unit 12 calculates the weight of the keyword assigned to the content from the relationship between each content and the keyword. Subsequently, in step S3, the content-related score calculation unit 13 calculates a related score between the contents based on the weight of the keyword of the content. Thereafter, in step S4, the recommended content determination unit 14 determines the recommended content based on the related score, and the content information output unit 15 outputs the recommended content set 30.

  As described above, in this embodiment, the content recommendation device 10 acquires a keyword set assigned to content when the user views the content (S1), and stores the content and keywords stored in the keyword storage device. The keyword is weighted by analyzing the relationship (S2). Next, scoring is performed for each content stored in the keyword storage device 100 based on the keyword weighting (S3). Finally, one or a plurality of contents having higher scores are selected and presented to the user (S4).

  The keyword weight value calculated once is stored in the keyword storage device 100, so that the calculation of the same keyword weight can be omitted from the next time. Further, instead of calculating the weight of each keyword whenever necessary, the weight may be calculated in advance for all content keywords and stored in the keyword storage device 100.

  In the keyword storage device 100, data about content sets and keywords assigned thereto are stored in a content information management table 101, a keyword information management table 102, a keyword / content relationship management table as shown in FIG. 3 or 4, for example. 103 in a table format. The keyword storage device 100 outputs data on content and keywords assigned to the content by specifying a search condition for selecting the content. That is, the keyword weighting unit 12 and the content-related score calculation unit 13 can acquire the content information and the keyword set attached to the content by specifying the content from the keyword storage device 100, and specify the keyword. Thus, the keyword information and the content set to which the keyword is assigned can be acquired.

  The example of FIG. 3 is an example of information stored in the keyword storage device 100 when [Keyword Weight Calculation Example 1] described later is used. The keyword storage device 100 stores, as the content information management table 101, content ID, content name, and description information for uniquely identifying each content to be managed. The keyword information management table 102 stores information such as a keyword ID for uniquely identifying each keyword and a keyword name. The keyword / content relationship management table 103 stores information on combinations of content IDs and keyword IDs, and keyword weight values for the contents.

  On the other hand, the example of FIG. 4 is an example of information stored in the keyword storage device 100 when [Keyword Weight Calculation Example 2] described later is used. Information similar to that shown in FIG. 3 is stored as the content information management table 101, but the keyword information management table 102 stores keyword weight values in addition to keyword IDs and keyword names for uniquely identifying each keyword. doing. The keyword / content relationship management table 103 stores correspondence information between each content ID and keyword ID.

  The content information management table 101, the keyword information management table 102, and the keyword / content relationship management table 103 shown in FIGS. 3 and 4 are examples, and attributes such as names and descriptions may not be stored. May be stored. Further, there may be no area for storing the weights of keywords that have already been calculated.

  Assuming that the user browses the content a, a content recommendation method will be described.

[Keyword weight calculation example 1]
First, an example of a first technique for calculating keyword weights will be described. In calculating the keyword weight, for example, it is conceivable to give a higher weight to a keyword having a strong association with more other keywords attached to the same content.

  In the keyword weighting calculation, first, the strength of the relationship between two arbitrary keywords in the keyword set assigned to the content a is calculated. FIG. 5 shows an example of calculating the strength of the relationship between keywords for keywords assigned to content according to an embodiment of the present invention. The strength of the relationship between keywords can be defined by, for example, a chi-square value for the appearance frequency of keywords. At this time, the strength R (A, B) of the relationship between keywords is given by the following equation.

Here, N is the total number of contents existing in the content set stored in the keyword storage device 100, and [A, B], [￣A, ￣B], [￣A, B], [A, ￣B] ] Is the number of contents to which both keywords A and B are assigned, the number of contents to which neither keyword A or B is assigned, the number of contents to which keyword B is assigned but not assigned keyword A, and keyword A is assigned. Keyword B is the number of contents not assigned. Note that “￣” in “￣A” is a symbol on A (the same applies to B).

  [A], [￣A], [B], and [￣B] are the number of contents to which the keyword A is assigned, the number of contents to which the keyword A is not assigned, and the contents to which the keyword B is assigned. The number of contents to which no keyword B is assigned.

  Next, keyword weighting is determined. The value of the weight W (a, A) of the keyword A assigned to the content a can be defined by, for example, an average value of the strength of the relationship between the keyword A and another keyword assigned to the content a. At this time, W (a, A) is given by the following equation.

However, K _a is the set containing all the keywords assigned to the content a, N _Ka is the total number of keywords contained in the K _a.

The above keyword weighting method is based on the following knowledge.
[Knowledge 1]: The content that is expressed redundantly by more keywords is the main content.

Many keywords will try to express particularly important content in the whole content. Furthermore, in this example, the relationship between the keyword and the content is obtained based on the following knowledge.
[Knowledge 2]: Keywords expressing the main contents are highly related to the contents.

  Therefore, by analyzing how much the contents of keywords overlap each other, it is possible to analyze how much the keywords express the main contents of the contents.

[Keyword weight calculation example 2]
Next, an example of a second method for calculating keyword weights will be described. In calculating the keyword weight, for example, when a keyword is assigned, the topic is often the same topic, that is, when the keyword can narrow down the topic, a higher weight may be given.

  In the keyword weighting calculation, first, for each keyword assigned to the content, a set of keywords co-occurring with each keyword in all the contents is acquired.

  FIG. 6 is a diagram illustrating an example of calculating the number of co-occurrence between keywords according to an embodiment of the present invention. In FIG. 6, a keyword A represents a keyword whose concreteness is to be calculated. All the contents to which the keyword A is assigned are acquired from the keyword storage device 100, and all the keywords assigned to the contents set are acquired. However, keyword A itself is excluded. The example of FIG. 6 indicates that the keywords co-occurring with the keyword A are the keyword B, the keyword C, and the keyword D. The number of times each keyword co-occurs with the keyword A is C (A, B), C (A, C), and C (A, D), respectively. That is, in the keyword storage device 100, the number of contents to which both the keyword A and the keyword B are assigned is C (A, B).

Subsequently, in order to examine the deviation of the number of co-occurrence of keywords co-occurring with the keyword A, a value of the following equation E (A) is calculated. Here, K ′ _A represents a set of keywords co-occurring with the keyword A, and N _K′A = Σ _k C (A, k) (where Σ _k represents the sum of k∈K ′ _A ). ). In the example of FIG. 6, K ′ _A is a set including a keyword B, a keyword C, and a keyword D.

  If the deviation of the number of co-occurrence of keywords co-occurring with keyword A is small, E (A) becomes large, and if the deviation is large, E (A) becomes small. In other words, the smaller E (A) is, the more specific keyword it is.

  Finally, the keyword weight can be calculated by calculating the value of the following equation W (A). The value of W (A) increases as the keyword A is more specific.

However, σ is a parameter, and an arbitrary value of σ> 1 can be used.

  The above keyword weighting method is based on the following knowledge. Keywords are assigned reflecting the content. For example, the keyword “sports” is given to content related to tennis, baseball, soccer, and the like. On the other hand, the content to which the keyword “World Cup” is given is mostly related to soccer. This difference arises because the topic area represented by the keyword “sports” and the keyword “world cup” is different.

  Therefore, we focus on the keyword bias that co-occurs with each keyword. Now, since many contents to which the keyword “World Cup” is assigned are related to soccer, it is considered that many keywords assigned to the contents are also related to soccer. For this reason, the bias of keywords that co-occur with the “World Cup” increases. Conversely, since the keyword “sports” is assigned to content of various topics, the bias of co-occurring keywords is reduced. Therefore, it is possible to determine the concreteness of the keyword by analyzing the magnitude of the bias of the co-occurring keywords.

  An example of the keyword concreteness calculation is shown in FIG. The example of FIG. 7 is an example in which the concreteness is calculated for the keywords “sports” and “world cup”. Now, assume that keywords that co-occur with “sports” are “soccer”, “tennis”, and “baseball”, and the number of co-occurrence is 180, 120, and 200, respectively. At this time, when E (A) in the formula 1 is calculated, E (sports) = 0.468. On the other hand, the keywords that co-occur with “World Cup” are “soccer” and “meeting”, and the co-occurrence times are 90 times and 10 times, respectively. At this time, E (world cup) = 0.141. Therefore, since E (world cup) is smaller than E (sports), it can be understood that “world cup” is a more specific keyword than “sports”.

  After the keyword weighting, a related score is calculated for each content in the keyword storage device 100. For the calculation of the related score, for example, the total sum of the weights of keywords assigned to the contents can be used. In that case, the score of the content b is calculated by the following equation S (b).

However, the content a is content browsed by the user, K _a is a set of all keywords assigned to the content a, K _b is a set of all keywords assigned to the content b, and δ (KεK _b ) is a function that is “1” when the keyword k is assigned to the content b, and “0” otherwise. Further, when the second method for calculating the keyword weight is used as the keyword weighting method, it is assumed that W (a, k) = W (k).

  The recommended content is determined by calculating and comparing the related scores for all the content. When there is only one content recommended for the user, the content with the highest related score is presented, and when there are multiple content, the content is recommended by presenting some in order from the highest relevant score. Can do.

  In addition, the recommended content determination unit 14 can cluster content that is a candidate for recommendation by using a method of clustering a content set whose related score is equal to or greater than a threshold, and can recommend a wide range of content.

  For example, for each content that is a recommendation candidate, a related score is calculated using the formula [5]. Subsequently, the top n content items with the highest related scores are acquired. n may be an arbitrary value. Furthermore, by clustering these contents, it becomes possible to recommend contents with a wide range of contents.

  As a clustering method, for example, a vector having a weight for a keyword assigned to content as an element can be clustered by a k-means method (see Reference Document 1 below).

  [Reference 1]: J. McQueen, “Some methods for classification and analysis of multivariate observations”, In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, pp.281-297 (1967).

  Further, the user's browsing information is further accumulated, and the keyword weighting unit 12 calculates the weight based on the keyword information assigned to the plurality of contents by receiving the input of the plurality of contents browsed by the user. Can also be used. By using this method, it is possible to recommend content based on information on a plurality of contents viewed by the user.

  FIG. 8 is a diagram illustrating a configuration example of the content recommendation device 10 in the case where content recommendation is performed based on a plurality of pieces of content information browsed by a user.

  The content recommendation device 10 includes a user browsing information storage device 110 that stores user browsing history information. The user information input unit 16 receives recommended user information, acquires content information viewed by the user from the user browsing information storage device 110, and transmits the information to the keyword weighting unit 12. The keyword weighting unit 12 weights the keywords assigned to all the contents viewed by the user. The content-related score calculation unit 13 calculates a content-related score based on the keyword weight, and the recommended content determination unit 14 outputs the content with a high related score as content to be recommended by the content information output unit 15.

  For example, for each content viewed by the user, the weight of each keyword assigned to the content can be calculated, and the sum of them can be used to calculate the related score of the content. If the set of contents browsed by the user is C, the score of the contents b is calculated by the following equation S ′ (b).

  Furthermore, a means for changing the weight of the keyword by the user can be provided, and the keyword weight can be changed by selecting a keyword in which the user is interested or not interested in the keyword weighting unit 12. By using this method, when the recommendation result is not satisfactory, the user can arbitrarily change the weight of the keyword and change the recommended content.

  FIG. 9 is a diagram illustrating a configuration example of the content recommendation device 10 that can change the recommended content by a user operation.

  In this example, the user can increase or decrease the keyword weight by inputting an interesting keyword or an uninterested keyword to the keyword weight changing unit 17. After the user's operation, the keyword weight changing unit 17 transmits the contents of the user operation to the keyword weighting unit 12, and the keyword weighting unit 12 corrects the keyword weight. The content-related score calculation unit 13 calculates a content-related score based on the keyword weight, and the recommended content determination unit 14 outputs the content with a high related score as content to be recommended by the content information output unit 15.

  For example, for a user who browses a certain content, using the content recommendation device 10, the recommended content is presented, and at the same time, a set of keywords assigned to the content is presented to the user. When the user wants to change the recommended content, when the user selects an interesting keyword from the keyword set, the content recommendation device 10 increases the weight of the keyword. Or, if you select a keyword that you are not interested in, reduce the weight of the keyword. Then, based on the changed keyword weight, the content score is recalculated, the recommended content is determined, and presented to the user.

  Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and modifications and applications are possible within the technical scope described in the claims.

DESCRIPTION OF SYMBOLS 10 Content recommendation apparatus 11 Content information input part 12 Keyword weighting part 13 Content related score calculation part 14 Recommended content determination part 15 Content information output part 16 User information input part 17 Keyword weight change part 100 Keyword storage apparatus 101 Content information management table 102 Keyword Information management table 103 Keyword / content relation management table 110 User browsing information storage device

Claims (9)

A content recommendation device for recommending content to a user,
Content information input means for acquiring information of content viewed by the user;
A keyword storage means for storing keywords assigned to content;
A keyword weighting means for weighting a keyword based on at least one of a keyword-content relevance and a keyword concreteness;
Content that calculates the sum of the weights of keywords that are commonly included in both the content viewed by the user and each content in the recommended candidate content set as a related score between the content viewed by the user and each content A related score calculation means;
Recommended content determining means for determining one or a plurality of content having a large relevance score with the content viewed by the user as recommended content;
A content recommendation device comprising: content information output means for presenting information of the recommended content to a user. The keyword storage means stores information on each content to be recommended, information on keywords assigned to the content, and information on each keyword,
The keyword weighting means and the content-related score calculating means acquire content information and a keyword set attached to the content by designating the content from the keyword storage means, and designate the keyword to designate the keyword. The content recommendation apparatus according to claim 1, wherein a content set to which information and a keyword are assigned is acquired. The keyword weighting means is:
Calculate the chi-square value of the appearance frequency of any two keywords assigned to the content, and calculate the average value of the chi-square value of each keyword and other keywords, thereby calculating the relevance between the content and the keyword The content recommendation device according to claim 1, wherein a value indicating the calculated degree of relevance is used as a weight of the keyword. The keyword weighting means is:
For all the contents, the keyword that co-occurs with the keyword to be weighted is obtained, and the value indicating the concreteness of the keyword is calculated by calculating the bias of the co-occurrence frequency of the co-occurring keyword. The content recommendation device according to claim 1, wherein a value indicating sex is used as a weight of the keyword. The recommended content determination means includes:
The content recommendation apparatus according to any one of claims 1 to 4, wherein clustering of a content set having a related score equal to or greater than a threshold value is performed. Furthermore, it has user browsing information storage means for storing the user's content browsing information,
The keyword weighting means is:
The keyword weight is calculated based on the keyword information assigned to the plurality of contents by referring to the user browsing information storage means and accepting input of the plurality of contents browsed by the user. The content recommendation device according to any one of claims 1 to 5. In addition, there is a keyword weight changing means for selecting a keyword that the user is interested in or not interested by input from the user, and changing the weight of the selected keyword,
The content recommendation device according to any one of claims 1 to 6, wherein the keyword weight calculated by the keyword weighting means can be changed.   The content recommendation program for functioning a computer as a means with which the content recommendation apparatus of any one of Claim 1-7 is equipped.   A computer-readable recording medium on which the content recommendation program according to claim 8 is recorded.

Images