JP5373743B2 - Similar user extraction method, similar user extraction device, and similar user extraction program - Google Patents

Description

  The present invention relates to a technique for extracting similar users of emphasis filtering that recommends items possessed by similar users based on user ratings.

  Currently, based on the user's score, recommend similar items based on the similarity of attributes such as emphasis filtering that recommends items of similar users with similar rating actions to the recommended user and metadata associated with the items. Content filtering is used by many content providers.

For example, Patent Document 1 proposes a user similarity based on a taxonomy (classification system) and a recommendation method based on the similarity. Here, using taxonomies that classify descriptive words of blog information of multiple users acquired from a blog site, the correlation coefficient between evaluation vectors based on evaluation values given to each class or item of the taxonomy is calculated as the similarity between users It is said. This correlation coefficient is calculated by dividing the covariance of two evaluation value vectors by the product of the standard deviation of each vector. For example, the evaluation value vector of each user is determined from the evaluation values of user A and user B. And these correlation coefficients (gamma) _AB are calculated like Formula (1). In formula (1), σ _A. σ _B indicates a standard deviation, and A ′. B ′ is the user A.B. B shows the average of the evaluation values given to each class or item of taxonomy, and Cov (A, B) shows covariance.

  Further, in Non-Patent Document 1, in the recommendation technique using the topic model, considering the order of purchase (time series), to the recommended user based on the items consumed by the user who purchased the product in the same time series. The recommendation is determined.

JP2008-117122A

Tomoharu Iwata, Shinji Watanabe, Takeshi Yamada, Naonori Ueda, “Topic Tracking Model for Analyzing Purchasing Behavior”, Proc. of 21st International Joint Conference on Artificial Intelligence (IJCAI-09), 1427-1432, 2009

  However, conventionally, no attempt has been made to extract not only a rating action but also a similar user in consideration of the context (motivation, sensitivity, restrictions, etc.) when the user accesses the item.

  That is, since the taxonomy in the method of Patent Document 1 can handle only one axis, for example, in addition to the genre, if the user similarity is calculated in consideration of the user's context based on another axis such as time, place, and sensitivity. As shown in FIG. 1, it is necessary to construct one taxonomy in which both are mixed.

In this case, since the user u is interested in both the item I ₁ under the genre C ₂ and the I ₁ under the sensitivity C ₄ , the user u's interest in I ₁ is calculated larger than necessary. End up. This makes it impossible to accurately calculate the similarity between users, and it is impossible to extract similar users in consideration of the context.

  In addition, the recommendation technique using the topic model of Non-Patent Document 1 can be applied only to time series, and since taxonomies are not accepted as input, data that is valuable by systematically organizing sensibilities and places (Context) cannot be processed.

  The present invention has been made in order to solve the above-described problems, and an object of the present invention is to extract similar users in consideration of the context in collaborative filtering.

Normally, as shown in FIG. 2, a user accesses an item according to some context (motivation, sensitivity, restriction). Here, the user u accesses the item I 1 through the context ₁ , and the user v accesses the item 1 through the context 2.

In FIG. 3, the user u feels that the item I ₁ is cool and accesses the user I, while the user v feels that the item I ₁ is surprisingly accessed. Here, even if they are in contact with the same item, each user feels differently about the item. Also, in FIG. 4, user u accesses item I ₁ in Shibuya, while user v accesses item I ₁ in Kyoto, and the location for the item is different.

  The present invention pays attention to such a context, and uses a two-dimensional similarity such as (1) the similarity of context between users when accessing an item, and (2) the similarity of evaluation degree of items between users. To calculate the user similarity. In calculating the context similarity, it must be considered that the contexts between users may be similar but may not be exactly the same. For example, as shown in FIG. 5, an item may be accessed through a geographically similar context of “Omotesando” and “Daikanyama”.

  Therefore, in the present invention, a context classification system, that is, a context taxonomy, is used to calculate the similarity in consideration of subtle differences in context between users. 6 and 7 show an example of the context taxonomy, FIG. 6 shows the context taxonomy of the classification system “location”, and FIG. 7 shows the context taxonomy of the classification system “mod”. The similarity of the context calculated based on this taxonomy is reflected in the score given to the user's item, and the similarity of the score evaluation between users, that is, the similarity between users is obtained.

  Specifically, the recommended information detection / providing method of the present invention is based on the context taxonomy of the context that is the basis of access to the item and the item taxonomy of the item that is the item classification system. Context similarity calculation step to calculate the similarity of the context, and the score to obtain the newly updated score by reflecting the similarity of the context calculated in the context similarity calculation step to the score given to each item by each user A reflecting step, and a similar user extracting step of extracting similar users of the recommended users according to the similarity between users based on the scores updated in the rating reflecting step.

  In addition, the recommended information detection and provision apparatus of the present invention provides a context taxonomy of contexts obtained by classifying contexts as a basis for access to an item, and an item taxonomy of items obtained by classifying items into a classification system. Context similarity calculating means for calculating similarity, and score reflecting means for reflecting the similarity of the context calculated by the context similarity calculating means to the score given to each item by each user to obtain a newly updated score And similar user extracting means for extracting similar users of the recommended users according to the similarity between users based on the scores updated by the score reflecting means.

  In addition, this invention is good also as an aspect of the program which makes a computer function as said apparatus. This program can be provided through a network or a recording medium.

  According to the present invention, it is possible to extract similar users in consideration of the context in collaborative filtering. Therefore, it is possible to recommend an item of a user having a similar context to the recommended user.

The image figure which built the taxonomy of genre and sensitivity in patent documents 1. The image figure of a user's item access. The image figure from which the context (mode) with respect to the item differs. The image figure where the context (location) for the item is different. The image figure which shows the state which is the same geographically but is not the same. The image figure using a classification system (location). The image figure using a classification system (mode). The image figure which shows the relationship between a general user, an item, and context. The block diagram which shows the apparatus structural example of the similar user extraction apparatus which concerns on embodiment of this invention. The flowchart which shows the basic process step of the same user extraction device. The image figure of the state in which the same users access the same item. FIG. 12 is a development view of the context of FIG. 11. The image figure which reflects context similarity in the similarity of a score. The image figure which shows the process example of the similar user extraction part based on a context. The image figure which shows the state which is accessing the class of the same item instead of the same item between users. FIG. 16 is a development view of the context of FIG. 15. The image figure which accesses the item from which an item class differs between users by the same context class. The development view of the context of FIG.

  The similar user extraction apparatus according to the embodiment of the present invention extracts not only a rating action but also a user who is similar to a context as a basis for access to an item when recommending an item possessed by a user similar to the recommended user.

That is, when accessing the item, as shown in FIG. 8, the user accesses the item along some context. Here, the user u accesses the item I ₁ through the context C ₃ ^C , and the user v accesses the item I ₂ through the context C ₄ ^C. Therefore, the apparatus uses user interest information that expresses the user's interest by the context taxonomy, the item taxonomy, and the user's rating given to the item, and the similarity of the interest information is the similarity between users.

≪Example of device configuration≫
A configuration example of the apparatus will be described with reference to FIG. This apparatus 1 is incorporated in an emphasis filtering system such as a recommendation information detection apparatus disclosed in Patent Document 1, for example, and is a normal computer hardware resource such as a CPU. Memory (RAM). A hard disk drive device is provided. As a result of the cooperation between the hardware resource and the software resource, the apparatus 1 has a similarity calculation unit 2 based on the context 2, a score reflection unit 3 based on the context 3, and a similarity calculation unit 4 based on the taxonomy. The similar user extraction unit 5 based on the context is implemented.

  The similarity calculation unit 2 includes a context taxonomy (classification system) such as sensitivity, time, and place, an item taxonomy (classification system), and a user ID / consumed item / context / score. Access information is input. The score in the access information indicates a value (for example, five-level evaluation value) given to the item when the user accesses the item (item purchase or the like) with a user terminal (not shown).

  As shown in Fig. 8, the context taxonomy consists of a hierarchical structure in which the context is defined by some sort of classification system (sensitivity, time, place, etc.), and the item taxonomy consists of a hierarchical structure that classifies the concept of the item. Are stored in databases not shown, and both taxonomies are input to the similarity calculation unit 2 with reference to the respective databases. This database may be built in a memory / hard disk drive device built in the device 1 or may be built in an external input device, and each taxonomy is prepared by a content provider. In addition, what was created from the blog information of patent document 1 can also be used for each said taxonomy.

  Specifically, the similarity calculation unit 2 calculates a context similarity when an item is accessed between users having a user ID based on the input data. Cases of this similarity calculation include (1) a simple case of calculating the context similarity between users when the same item is consumed, and (2) a class position (node position) on the hierarchical structure of the item taxonomy. Extended case 1 to calculate the similarity of context between users when consuming the same class, (3) Utilizing the parent-child relationship in the hierarchical structure of the context, even if the context is not the same, the parent node is the same context If so, there is an extended case 2 for calculating the degree of context similarity between users. The calculated similarity / access information is sent to the score reflecting unit 3.

  The score reflection unit 3 receives the similarity calculated by the similarity calculation unit 2 and the access information. Here, in order to calculate the evaluation degree of the item between users, the similarity of the context is reflected in each user's score, and the score is updated. The access information including the updated score is sent to the similarity calculation unit 4.

  Access information including the score updated by the score reflection unit 3 is input to the similarity calculation unit 4 and the taxonomy of the item is input again. Here, the updated score is set as an evaluation value given to each class or item in Patent Document 1, and a correlation coefficient between evaluation value vectors is calculated as the closeness of user interest information, that is, the similarity between users. The calculated similarity between users and access information are sent to the similar user extraction unit 5.

  The similar user extraction unit 5 receives the similarity between users calculated by the similarity calculation unit 4 and the access information, as well as the context information given by the recommended user. This context information is a context specified by the recommended user through the user terminal. For example, a context information input field may be displayed on the user terminal of the recommended user, and the context input by the recommended user to the input field and transmitted to the device 1 may be used.

  Here, a recommended candidate user group having a high degree of similarity between the recommended user and the user is selected, and recommended candidate users having the context indicated by the context information in the access information are extracted as a set of similar users. Since the extracted similar user set is sent to the recommendation processing unit of Patent Document 1, it is possible to recommend similar user items reflecting the recent context to the recommended user.

≪Process step≫
The processing steps of the device 1 corresponding to the simple case will be described with reference to FIG. Here, the context similarity is calculated only when the users are accessing the same item in the taxonomy of the item. That is, even if the users access the same item, if the contexts are different, the user similarity is calculated considering the difference.

  Note that the context classification system interposed between the user and the item is not necessarily one type, and a plurality of context classification systems (for example, a combination of sensitivity and time, etc.) are generally used. Then, it explains as one kind of classification system.

S01. S02: The similarity calculation unit 4 performs S01. Context similarity between users is calculated through S02. An example of this calculation will be described with reference to FIGS. Here, the context similarity with respect to the item I ₁ and the context similarity with respect to the item I ₂ are calculated separately. In this processing example, calculation of the similarity of the context with respect to the item I ₁ will be described.

First, the similarity calculation unit 4 uses Equation (2) as a vector expressing whether the user u is accessing an item via a subclass under the class C _i ^C in the context taxonomy, Using the vector as expression (3), the value of each element is extracted (S01).

At this time, the value of each element of the vectors of Expressions (2) and (3) is “1” if the item is accessed via the corresponding subclass under the subordinate, and “0” if not accessed. . The values of the elements in the vectors of equations (2) and (3) are determined for all classes “C _i εC” in the context taxonomy.

In FIG. v is the context C ₃ ^C. Item I ₁ is accessed by C ₄ ^C. Here, as shown in FIG. 12, Kontekuto C ₃ ^C. C ₄ ^C belongs to a subclass of class C ₁ ^C. The value of the vector expression for the item I ₁ in this case (4) is determined to be (1,0), the values of the elements of the vector of the formula (5) for the items I ₁ is determined to (0,1). The same applies to other classes in the context.

Next, the similarity calculation unit 4 receives the user u. The context similarity for the item I ₁ between v is calculated using the cosine similarity (S02). That is, the cosine similarity is calculated for the element values of each class of the equations (2) and (3) obtained in S01.

User u on the final item I ₁ by summing the result of this calculation. Context similarity between v is obtained. For example, in FIG. 12, the user u for items I _1. The context similarity S (I ₁ ) _{u, v} between _v is calculated as in equation (6).

S03: The score reflection unit 3 reflects the similarity of the context when the item is accessed when calculating the similarity of the evaluation degree of the item between the users, that is, the similarity of the score. For example, user u. In order to calculate the score similarity between v, as shown in FIG. 13, the score (2.0) of the user u for the item I ₁ is the context similarity S (I ₁ ) _u calculated in S02. _{, v} . Similarly, the score (4.0) of the user v for the item I ₁ is multiplied by the context similarity S (I ₁ ) _{u, v} calculated in S02. As a result, the user u. The rating value for v item I ₁ is updated.

  S04: The similarity calculation unit 4 determines the user u. Based on the score value updated in S03. The similarity between v is calculated. In the calculation, the updated score value is set as the evaluation value of Patent Document 1, and the correlation coefficient between the evaluation value vectors based on the evaluation value is set as the user u. Calculated as the similarity between v. Here, the updated score value is substituted into equation (1) of Patent Document 1, and the user u. What is necessary is just to calculate the similarity of v.

  S05: The similar user extraction unit selects a recommended user group whose context information is input and a recommended candidate user group whose similarity between users calculated in S04 is equal to or greater than a threshold value. The selected recommendation candidate users are further narrowed down to recommendation candidate users having a context indicated by the context information in the access information. The refined recommended candidate user group is extracted as a final similar user set.

If an example is demonstrated based on FIG. 14, CT1 will show the context information of the recommended user u, and CT2 will show the access information of the recommendation candidate user v. Here, it is assumed that the context information C ₄ ^C exists in the access information of the user u, but the context information CT 2 does not include the context C ₄ ^C.

In the access information CT2, there is only access to an item based on the context C ₄ ^C , and the context C ₃ ^C. Since it does not have C ₅ ^C , user v is not included in the similar user set of user u.

Therefore, according to the device 1, a similar user corresponding to the context information input by the recommended user is extracted. Thereby, the recommendation information according to the context of the recommended user that changes with time can be preferentially presented through the recommendation processing unit of Patent Document 1. For example, with respect to favorite user actor that recently C ₃ ^c, a user with the context of actor C ₃ ^c, it is possible to recommend an item to have a closer evaluation of the item the user.

  That is, by dividing the context taxonomy and the item taxonomy and constraining the context usage range according to time, it is possible to recommend items according to the recent context. At this time, not only the context axis of the actor but also the context axis such as genre and sensitivity can be used to derive the relevance of the recommendation based on various context axes.

≪Expansion case 1 processing≫
The context similarity calculation according to the expansion case 1 will be described below. In this case, if the user accesses an item belonging to the same class instead of the same item, S01. Through S02, the similarity of the context for the class is calculated. This is because the difference in context when users are accessing items of the same class is reflected in the user similarity and the recommendation accuracy is improved.

Referring to FIG. 15, item I ₁ . I ₂ is classified under the same class C ₁ ^T. In this case, user u is accessing item I ₂ with context C ₃ ^C , while user v is accessing item I ₁ with context C ₄ ^C. Here, both users u. Each item I ₁ . Substituting access to I ₂ with access to class C ₁ ^T , the vectors of equations (4) and (5) are obtained (S01).

In this case, as shown in FIG. 16, the item I _1. The vector of equation (4) for I ₂ is determined as (1, 0), and the vector of equation (5) is determined as (0, 1). Other processing contents are S01. The same as S02.

≪Expansion case 2 processing≫
The context similarity calculation according to the extension case 2 will be described below. Here, if the user uses the same class on the context taxonomy, the context similarity is reflected in the similarity of the item evaluation action even if the class (or item) on the item taxonomy is different.

  This is because even if items are under different classes, it is considered that the interests of the user are correlated if they are in the same context class. For example, if an item is a movie, an item taxonomy is categorized by genre, and a context taxonomy is categorized by actor, if the genre is different but the actors are the same, the user interest may be correlated. It is done.

Referring to FIG. 17, the user u accesses the item I ₃ using the context C ₃ ^C , while the user v accesses the item I ₁ using the context C ₃ ^C. Here, user u. v is, the use of the same context C ₃ ^C, as shown in FIG. 18, the item I _1. The vectors of equations (4) and (5) for I ₃ are both determined as (1, 0). Other processing contents are S01. The same as S02.

  In addition, this invention is not limited to the said embodiment, It can deform | transform within the range described in each claim. For example, the similar user extraction unit 5 does not necessarily need to narrow down the recommended candidate users according to the context information input by the recommended user, and can also extract the recommended candidate users as a final similar user set. In this case, although the context information given by the recommended user is not reflected, a set of similar users reflecting the similarity based on the context taxonomy is extracted, so the item of the similar user is recommended to the recommended user Can do.

≪Programs≫
The present invention can also be configured as a document search program that causes a computer to function as some or all of the units 2 to 5 of the apparatus 1. According to this program, it is possible to cause the computer to execute part or all of S01 to S05.

  The program can be provided through a network such as a website or e-mail. The program is stored in a recording medium such as a CD-ROM, DVD-ROM, CD-R, CD-RW, DVD-R, DVD-RW, MO, HDD, BD-ROM, BD-R, or BD-RE. It is also possible to record, save and distribute. This recording medium is read using a recording medium driving device, and the program code itself realizes the processing of the above embodiment, so that the recording medium also constitutes the present invention.

DESCRIPTION OF SYMBOLS 1 ... Similar user extraction apparatus 2 ... Context based similarity calculation part (Context similarity calculation means)
3… Score reflection unit based on context (score reflection means)
4 ... Similarity calculation unit based on taxonomy 5 ... Similar user extraction unit based on context (similar user extraction means)

Claims (9)

A method of executing an apparatus for extracting similar users in recommending items to similar users who have similar scoring actions based on user scores,
A context similarity calculation step for calculating a context similarity between users for an item based on a taxonomy of a context in which the context that provides access to the item is classified and a taxonomy of the item in which the item is classified
A score reflection step for obtaining a newly updated score by reflecting the similarity of the context calculated in the context similarity calculation step in the score given to each item by each user;
A similar user extracting step of extracting similar users of the recommended users according to the similarity between users based on the respective scores updated in the rating reflecting step;
A similar user extraction method characterized by comprising: The similar user extraction method according to claim 1, wherein the context similarity calculation step calculates a context similarity between users accessing the same item or the same class item in the taxonomy of the item. The similar user extraction method according to claim 1, wherein the context similarity calculation step calculates a context similarity between users who access the item in the same class of context in the context taxonomy. The similar user extraction step according to any one of claims 1 to 3, wherein the similar user extraction step includes extracting the similar users only to users having contents designated by the recommended user. An apparatus for extracting similar users when recommending to a recommended user an item possessed by similar users having similar scoring actions based on the user's scores,
A context similarity calculation means for calculating the similarity of contexts between users for an item based on the taxonomy of the context in which the context that provides access to the item is classified and the taxonomy of the item in which the item is classified
A score reflecting means for obtaining a newly updated score by reflecting the similarity of the context calculated by the context similarity calculating means in the score given to the item by each user;
Similar user extraction means for extracting similar users of recommended users according to the similarity between users based on the scores updated by the score reflection means;
A similar user extraction device comprising: 6. The similar user extraction device according to claim 5, wherein the context similarity calculation means calculates a context similarity between users accessing the same item or the same class item in the taxonomy of the item. The similar user extraction device according to claim 5, wherein the context similarity calculation means calculates a context similarity between users who access the item in the same class of context in the context taxonomy. The similar user extraction unit according to any one of claims 5 to 7, wherein the similar user extraction unit extracts the similar users only to users having contents designated by the recommended user.   The similar user extraction program for functioning a computer as each means of the similar user extraction apparatus of any one of Claims 5-8.

Images