JP2011034379A - Cooperative filtering system, method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a cooperative filtering system which enables cooperative filtering considering asymmetry in the strength of relationship between pairs for calculating the relationship. <P>SOLUTION: A user action history DB 321 stores history of action applied to an item by a user. A user pair relationship calculation means 323 refers to the user action history DB 321 and calculates relationship between user pairs in which one-to-one asymmetry is allowed (for U2UCF). An item score calculation means 325 calculates an item score by using the relationship between the user pairs and the user's action history information. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a collaborative filtering system, method, and program, and more particularly to a collaborative filtering system, method, and program used for information recommendation technology.

  There is a technique called collaborative filtering (hereinafter also referred to as CF (Collaborative Filtering)). The CF includes a U2UCF that correlates between users (U2U: User to User) and an I2ICF that correlates between items and information (I2I: Item to Item). CF is practically used for content recommendation in Amazon, Netflix, and the like.

  4A and 4B show the principle in U2UCF and I2ICF. FIG. 4A shows the relationship between users and items in U2UCF. The user 11 has a history referring to the item 13. The user 12 has a history referring to the item 13 and the item 14. The CF inputs the history of the user 11 and the history of the user 12. For the item 13, the user 11 and the user 12 are common, and a correlation between the users is obtained. Therefore, the hypothesis that the user 11 and the user 12 are users having similar preferences is established. Therefore, the item 14 having a history only for the user 12 is recommended to the user 11 on the assumption that the correlated user 11 is also preferred.

  FIG. 4B shows the relationship between users and items in I2ICF. The user 11 has a history that refers to the item 13, and the user 12 has a history that refers to the item 13 and the item 14. The CF inputs the user 11 and the history of the user 12. For the user 12, the item 13 and the item 14 are common, and the item correlation is obtained. Therefore, the hypothesis that the item 13 and the item 14 are similar to each other holds. Therefore, the item 14 is recommended to the user 11 on the assumption that the item 14 is also favored by the user 11 who has a history only in the item 13 correlated with the item 14.

  In CF, a correlation (relationship) between users or items is calculated. For the correlation calculation, a Pearson correlation coefficient, a Spearman rank correlation coefficient, or the like is used. Hereinafter, Pearson will be described as an example of the relationship calculation method. In the following, in order to avoid redundancy, the relationship calculation in the U2UCF will be described, but the relationship can also be calculated in the same manner in the I2ICF.

A user pair whose correlation is to be calculated is user 1 and user 2. The user behavior history of the user 1 and X _1, the user behavior history of the user 2 and X _2. The action history X is represented by a vector having an action history performed by the user for each item as an element. The elements of the user behavior histories X ₁ and X ₂ are x ₁ (i) and x ₂ (i) (i = 1 to n). n is the total number of items. Elements x ₁ (i) and x ₂ (i) each represent a user's action history for the i-th item. Here, the values of the elements x ₁ (i) and x ₂ (i) are defined as 1 when the user includes the i-th item in the history, and 0 when not included. The value of the element x (i) is not limited to a value indicating whether or not the element is referenced, but may be another value indicating an action on the item such as an evaluation performed by the user on the item or a time when the item is referred to.

The correlation C (X ₁ , X ₂ ) between user pairs is obtained by the following equation.

Xiaodan Song, Yun Chi, Koji Hino, Belle L. Tseng, “Identifying Opinion Leaders in the Blogosphere”, CIKM ’07, November 2007.

  In U2UCF, the relationship between users is derived from the common part between user pairs based on the item reference history etc. performed by the user, using the size of the common term and the closeness of the evaluation content made for the item as an index. is doing. The magnitude of this relationship is the same in both directions. That is, the strength of the relationship of the user Y viewed from the user X is equal to the strength of the relationship of the user X viewed from the user Y. However, in general, human relationships often do not have the same strength. In collaborative filtering, since the correlation between pairs is calculated with the same strength in both directions, collaborative filtering that reflects the actual relationship between pairs cannot be performed.

  Regarding the relationship between users, Non-Patent Document 1 places importance on a person who has left an early history as an Opinion Leader. However, this is not adapted for CF. Non-Patent Document 1 focuses only on temporal relationships. For example, when there are three users A, B, and C in the order of early action, the importance is “A> B”. , “B> C”. Therefore, “A> C”. However, in practice, relationships such as “A> B”, “B> C”, and “A <C” are also conceivable. Non-Patent Document 1 cannot cope with such a relationship. In Non-Patent Document 1, the relationship is only a specific example in terms of time, and the amount of the user's behavior history and the ratio of a certain preference history in the user's entire behavior history are taken into account. It was not what I did.

  It is an object of the present invention to provide a collaborative filtering system, method, and program that enable collaborative filtering in consideration of the asymmetry of the strength of the relationship between pairs for calculating the relationship.

  In order to achieve the above object, the present invention refers to a user action history DB that stores history information of actions performed by a user on an item, and allows one-to-one asymmetry to be allowed. Collaborative filtering comprising relationship calculation means for calculating the relationship between pairs or context pairs, and item score calculation means for calculating an item score using the relationship between the pairs and the action history information Provide a system.

  The present invention refers to a user behavior history DB in which a computer stores history information of behavior performed by an user on an item, and a one-to-one asymmetry is allowed, between user pairs, between item pairs, or There is provided a collaborative filtering method including a step of calculating a relationship between context pairs, and a step of calculating an item score using the relationship between the pair and the action history information.

  The present invention refers to a user action history DB that stores history information of actions performed by a user on an item in a computer, and allows one-to-one asymmetry, between user pairs, between item pairs, or A program for executing a process for calculating a relationship between context pairs and a process for calculating an item score using the relationship between the pairs and the action history information is provided.

  The collaborative filtering system, method, and program of the present invention can perform collaborative filtering in consideration of the asymmetry of the strength of the relationship between the pairs for calculating the relationship.

The block diagram which shows the collaborative filtering system of 1st Embodiment of this invention. The flowchart which shows an operation | movement procedure. The figure which shows the example of data of user pair relationship preservation | save DB. (a) And (b) is a block diagram which shows the principle of U2UCF and I2ICF, respectively.

Prior to the description of the embodiment, the significance of giving asymmetry to the relationship will be described. As a user pair, two users (user 1 and user 2) having a significantly different number of referenced items are considered. That is, when the total number Σx ₁ (i) of user 1's action history and the action history Σx ₂ (i) of user 2 are compared,
Σx ₁ (i) << Σx ₂ (i)
Suppose that

となる。ユーザ１が参照し、ユーザ２は参照しなかったアイテムの数は、

となる。ユーザ１は参照せず、ユーザ２が参照したアイテムの数は、

となる。ユーザ１もユーザ２も参照しなかったアイテムの数は、

となる。 Assume that the total number of action histories of user 1 is 110, the total number of action histories of user 2 is 10100, and the total number of common parts of the action histories is 100. The number of items commonly referred to by user 1 and user 2 is n as the total number of items.

It becomes. The number of items that user 1 references and user 2 does not reference is:

It becomes. The number of items that user 1 did not refer to and user 2 referred to is

It becomes. The number of items that neither user 1 nor user 2 referred to is

It becomes.

  Since there are 100 items commonly referred to by user 1 and user 2, user 1 and user 2 have a correlation. However, when the total number of items referred to by the user 2 is compared with the number of items commonly referred to, the number of items commonly referred to is smaller than the total number of items. In other words, if the user 2 has a large number of reference histories for items for which the user 1 does not leave a reference history, and the user 2 is a user who has a reference history for various items, the user 2 It is thought that 100 items out of the various items referred to only coincide with the items that the user 1 referred to. Therefore, the remaining 10,000 items for which the user 2 has a reference history are not used as a reference for the user 1 and should not be recommended to the user 1.

  On the other hand, the total number of items referred to by the user 1 is 110, and the number of items referred to in common with the user 2 is 100. Therefore, the ratio of the commonly referred items in the total number of items referred to by the user 1 is high. Therefore, for the user 2, the items that the user 1 refers to and the items that the user 1 does not refer to should be helpful, and the remaining 10 items should be recommended to the user 2. Thus, the user 2 is not helpful for the user 1, but the user 1 is actually helpful for the user 2. However, such a case cannot be handled when the correlation between user pairs has symmetry.

  Therefore, in the present invention, a relationship calculation unit that refers to the user behavior history DB and calculates a relationship in which one-to-one asymmetry is allowed between user pairs, item pairs, or context pairs is used. A minimum configuration of the present invention is obtained by adding an item score calculating means for calculating an item score using the relationship between the pair and the action history information stored in the user action history DB to the relationship calculating means. . The relationship calculation means calculates the relationship in which asymmetry is allowed between the pairs, and calculates the item score using the relationship, taking into account the asymmetry of the strength of the relationship between the pairs. Collaborative filtering is possible. For example, in the above case, the user 2 viewed from the user 1 is weakened, and the user 1 viewed from the user 2 is strengthened so that the user 1 only recommends an item having a reference history to the user 2. However, it is possible to prevent the user 1 from recommending items that have a reference history only to the user 2.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 shows a collaborative filtering system according to a first embodiment of the present invention. The collaborative filtering system has a recommendation engine 30. The recommendation engine 30 is a computer system such as a server. The recommendation engine 30 is connected to the user terminals (21, 22, 23) via a network or the like. The recommendation engine 30 carries out reception and holding of user behavior history, reception of user search requests, output of recommended items, score calculation, and the like. In FIG. 1, the number of user terminals is three, but is not limited thereto.

  The user terminals 21, 22, and 23 are terminal devices used by users such as mobile phones and PCs (Personal Computers). The user can operate the user terminals 21, 22, and 23 to request the recommendation engine 30 to recommend items (information). The recommendation engine 30 recommends items according to the request. The user terminals 21, 22, and 23 display recommended contents. In addition, the user can use the user terminals 21, 22, and 23 to transmit behavior information such as his / her item reference behavior to the recommendation engine 30.

  The recommendation engine 30 includes a user IF 31, a scoring unit 32, an item data storage DB 34, and a user data storage DB 35. The user IF 31 includes a user behavior input IF 311, a recommendation request reception IF 312, and a recommended item output IF 313. The scoring unit 32 includes a user behavior history DB 321, a weight calculation unit 322, a user pair relationship calculation unit 323, a user pair relationship storage DB 324, an item score calculation unit 325, and a correlation calculation unit 326.

  The user IF 31 is a part that transmits information between the user terminals 21, 22, and 23 and the scoring unit 32. The scoring unit 32 calculates an item score and returns the result to the user IF 31. The item data storage DB 34 stores information about items such as item names and IDs. The user data storage DB 35 stores information about the user such as the user name and ID. The item data storage DB 34 and the user data storage DB 35 do not have to be in the recommendation engine 30 and may be external databases. Further, the user behavior history DB 321 and the user pair relationship storage DB 324 of the scoring unit 32 do not need to be in the scoring unit 32 and may be an external database.

  The user behavior input IF 311 receives user behavior histories from the user terminals 21, 22, and 23. The user's action history includes information such as a reference history of items executed by the user. The user behavior input IF 311 stores the received user behavior history in the user behavior history DB 321. The user action history DB 321 includes data items such as user ID, item ID, action such as reference and bookmark, and date and time.

  The recommendation engine 30 may have action execution means for receiving an item reference request from a user. The action execution means includes item detail reference means, item detail data storage DB, item bookmark means, item bookmark DB, and the like. The item detail reference means receives an item detail reference request from the user terminal via the user IF 31. The item detail reference means stores detailed data of the referenced item in the item detail data storage DB. The item bookmark means receives an item bookmark request from the user terminal via the user IF 31. The item bookmark means stores the requested item in the item bookmark DB. In the case of such a configuration, the action execution means may store in the user behavior history DB 321 the behavior history that the user referred to the detailed data and the behavior history that the user bookmarked the item.

  The recommendation request reception IF 312 receives an item recommendation request transmitted by the user terminal. The user can specify the condition of the item for which recommendation is requested using, for example, a search keyword. When receiving the recommendation request, the recommendation request reception IF 312 transmits the received recommendation request to the item score calculation unit 325. The recommended item output IF 313 receives a list of recommended items and its score from the item score calculation means 325. The recommended item output IF 313 transmits the list of recommended items and the score thereof to the user terminal of the user who transmitted the item recommendation request.

  The user pair relationship calculation means 323 refers to the user behavior history DB 321 and calculates the relationship between user pairs that allows one-to-one asymmetry. The user pair relationship calculation unit 323 includes a weight calculation unit 322 and a correlation calculation unit 326. The weight calculation unit 322 refers to the user action history DB 321 and calculates a weight between user pairs. The weights calculated by the weight calculation unit 322 are allowed to be different in each user pair in both directions. That is, the weight of another user q viewed from a certain user p and the weight of the user p viewed from the user q are allowed to be different. For example, the weight calculation unit 322 determines the weight to the other viewed from one of the pairs based on the amount of behavior history existing only in the other of the pair. Alternatively, the weight calculation unit 322 may determine the weight based on the order of the time of the action history between the pairs.

  The correlation calculation unit 326 refers to the user action history DB 321 and calculates a correlation coefficient indicating the correlation between the user pairs. Any correlation coefficient can be used as the correlation coefficient. The correlation calculation unit 326 calculates, for example, a Pearson correlation coefficient. The user pair relationship calculation unit 323 calculates the relationship between user pairs from the weight between user pairs calculated by the weight calculation unit 322 and the correlation coefficient between user pairs calculated by the correlation calculation unit 326. The user pair relationship calculation unit 323 stores the calculated relationship between user pairs in the user pair relationship storage DB 324.

  The item score calculation means 325 refers to the user behavior history DB 321 and the user pair relationship storage DB 324, and calculates an item score based on the user behavior history and the relationship between the user pairs. The calculation of the item score can be performed by the same method as the calculation of the normal item score, except that the relationship between the user pairs is allowed to be asymmetric. The item score calculation means 325 transmits the calculated item score to the recommended item output IF 313.

The calculation of the weight will be described. When the user pair is a user p and a user q, the weight of the user q viewed from the user p is defined as w _pq . When the number of items commonly referred to by the user p and the user q in the total number of items referred to by the user q is large, that is, when the number of items referred only by the user q is small, the user q is helpful for the user p. It can be said that it is a user. Conversely, when the number of items commonly referred to by the user p and the user q in the total number of items referred to by the user q is small, that is, when the number of items referred only by the user q is large, the user q It can be said that the user p is not a very helpful user. Therefore, the weight w _pq is set so that the weight decreases as the number of items that do not overlap with the item referred to by the user p among the items referenced by the user q increases.

上記式１におけるｘ（ｉ）は、ユーザの行動履歴Ｘの要素であり、ユーザｘがｉ番目のアイテムを参照したときは１、参照していないときは０となる関数である。また、ｎは、アイテムの総数を表している。式１で定義される重みは、アイテムの総数をユーザｑのみが参照したアイテム数で割った値に相当する。ユーザｑから見たユーザｐの重みｗ_ｑｐは、式１にて、ｐとｑを入れ替えることで算出できる。重み算出手段３２２は、ユーザ行動履歴ＤＢ３２１からユーザｐ、ユーザｑの行動履歴を取得し、式１に基づいて、ユーザペア間の重みを算出する。 Specifically, the weight w _pq is defined by the following formula 1.

X (i) in the above equation 1 is an element of the user's behavior history X, and is a function that becomes 1 when the user x refers to the i-th item and 0 when it does not refer to it. N represents the total number of items. The weight defined by Equation 1 corresponds to a value obtained by dividing the total number of items by the number of items referenced only by the user q. The weight w _qp of the user p viewed from the user q can be calculated by exchanging p and q in Equation 1. The weight calculation unit 322 acquires the action history of the users p and q from the user action history DB 321, and calculates the weight between the user pairs based on Equation 1.

The correlation calculation unit 326 calculates a correlation coefficient between the user pairs from the user action history. Here, the correlation coefficient is a Pearson correlation coefficient. The correlation coefficient itself has symmetry. The user pair relationship calculating unit 323 calculates the product of the weight that the asymmetry calculated by the weight calculating unit 322 is allowed and the correlation coefficient having the symmetry calculated by the correlation calculating unit 326, and the relationship between the user pairs. And Specifically, the user pair relationship calculating unit 323 calculates the relationship of the user q viewed from the user p using the following formula 2.

  FIG. 2 shows an operation procedure. The user action history DB 321 stores user action history information. The weight calculation unit 322 refers to the user action history DB 321 and calculates the weight between the user pairs based on the action history information (step S1). Further, the correlation calculating unit 326 refers to the user action history DB 321 and calculates a correlation coefficient indicating the correlation between the user pairs based on the action history information (step S2).

  The user pair relationship calculation unit 323 stores the product of the weight and the correlation coefficient in the user pair relationship storage DB 324 as the relationship between the user pairs (step S3). For example, the user pair relationship calculation unit 323 calculates the relationship each time user behavior history information is stored in the user behavior history DB 321. Alternatively, the user pair relationship calculation unit 323 may calculate the relationship periodically at a predetermined time interval.

  The item score calculation means 325 receives a recommendation request via the recommendation request reception IF 312 (step S4). The item score calculation means 325 refers to the user behavior history DB 321 and the user pair relationship storage DB 324, and calculates an item score based on the user behavior history and the relationship between the user pairs (step S5). The item score calculation means 325 transmits the list of recommended items and the score to the request output source user terminal via the recommended item output IF 313 (step S6).

・ユーザ１が参照し、ユーザ２は参照しなかったアイテム

・ユーザ１は参照せず、ユーザ２が参照したアイテム

・ユーザ１もユーザ２も参照しなかったアイテム

Hereinafter, description will be made using specific numerical examples. The user of the user terminal 21 is user 1, the user of the user terminal 22 is user 2, and the user of the user terminal 23 is user 3. For the user pair (user 1, user 2), the number of items that the user 1 and the user 2 refer to in common from the action history X ₁ of the user ₁ and the action history X _{2 of the} user 2 stored in the user action history DB 321. Suppose that the number of items referred to only by the user 1, the number of items referred only by the user 2, and the number of items not referenced by both are as follows.
-Items that user 1 and user 2 have commonly referred to

Items that user 1 refers to, but not user 2

-Item that user 2 referred to without referring to user 1

・ Items that neither user 1 nor user 2 referred to

・ユーザ１が参照し、ユーザ３は参照しなかったアイテム

・ユーザ１は参照せず、ユーザ３が参照したアイテム

・ユーザ１もユーザ３も参照しなかったアイテム

In addition, regarding the user pair (user 1, user 3), the user 1 and the user 3 commonly referred to the user 1 from the action history X ₁ and the action history X _{3 of the} user 3 stored in the user action history DB 321. Assume that the number of items, the number of items referenced only by user 1, the number of items referenced only by user 3, and the number of items not referenced by both are as follows.
-Items that user 1 and user 3 commonly referred to

Items that user 1 refers to, but user 3 does not

・ Items referenced by user 3 but not by user 1

・ Items that neither user 1 nor user 3 referred to

・ユーザ２が参照し、ユーザ３は参照しなかったアイテム

・ユーザ２は参照せず、ユーザ３が参照したアイテム

・ユーザ２もユーザ３も参照しなかったアイテム

For the user pair (user 2, user 3), the number of items commonly referred to by the user 2 and the user 3 from the action history X ₂ of the user ₂ and the action history X _{3 of the} user 3 stored in the user action history DB 321 Suppose that the number of items that only the user 2 refers to, the number of items that only the user 3 refers to, and the number of items that both do not refer to are as follows.
-Items commonly referred to by user 2 and user 3

-Items referenced by user 2 but not by user 3

・ Items referenced by user 3 without referring to user 2

・ Items that neither user 2 nor user 3 referred to

In summary, the item reference status between user pairs is shown in Table 1 below.

The weight calculation means 322 calculates the bidirectional weight of each user pair using Equation 1. The total number n of items is assumed to be n = 100000. The weight between each user pair is
w ₁₂ = 10, w ₂₁ = 10000
w ₁₃ = 1000, w ₃₁ = 1000
w ₂₃ = 100,000 / 60, w ₃₂ = 100,000 / 10050
It becomes. Looking at the pair of users 1 and 2, the number of items that only the user 1 is referenced 10, the number of items that only the user 2 has reference is 10000, the weight w ₁₂ of the user 2 as seen from the user 1, the user The weight w _{21 of the} user 1 viewed from 2 is lighter. On the other hand, the pair of the users 1 and 3, since the number of items that only the number of items and the user 3 to only the user 1 is referenced referenced is the same number, has the same weight as the w ₁₃ and w _31.

であったとする。各ユーザペア間の関係性は、以下のように求まる。

The user pair relationship calculation means 323 calculates the relationship between user pairs using Equation 2. The average of the action histories of users 1 to 3 is

Suppose that The relationship between each user pair is obtained as follows.

  The user pair relationship calculation unit 323 stores the obtained relationship between user pairs in the user pair relationship storage DB 324. FIG. 3 shows relationship data stored in the user pair relationship storage DB 324. The relationship between user pairs can be represented by a square matrix in which users are arranged in rows and columns. However, there is no diagonal component value.

なお、推薦を要求するアイテムの条件が指定されている場合は、アイテムスコア算出に先立って、その条件に従って、条件に合致しないアイテムを除外しておいてもよい。その場合、スコア算出対象となるアイテムについて、アイテムスコアを算出する。 For example, it is assumed that the user 1 issues a recommendation request to the recommendation request reception IF 312 using the user terminal 21. The recommendation request reception IF 312 instructs the item score calculation unit 325 to calculate an item score. The item score calculation means 325 refers to the user action history DB 321 and the user pair relationship storage DB 324 to calculate an item score. Specifically, the item score calculation means 325 calculates an item score according to the following formula 3.

In addition, when the condition of the item for which recommendation is requested is specified, an item that does not meet the condition may be excluded according to the condition before calculating the item score. In that case, an item score is calculated for the item to be scored.

アイテムスコア算出手段３２５は、式３で求めたアイテムスコアのベクトルと、Ｉ−Ｘ_１ ^Ｔとをかけたものを求める。

上記演算を行うことで、ユーザ１がまだ参照していないアイテムにアイテムスコアが残る。アイテムスコア算出手段３２５は、式４で求めたアイテムスコアの中からスコアが大きいアイテムのアイテム名とスコアとを推薦アイテム出力ＩＦ３１３に渡す。 The vertical vector I is defined as follows.

The item score calculation means 325 obtains a product of the item score vector obtained by Expression 3 and I-X ₁ ^T.

By performing the above calculation, an item score remains on an item that the user 1 has not yet referred to. The item score calculation means 325 passes the item name and score of an item having a higher score from the item scores obtained by Expression 4 to the recommended item output IF 313.

  The recommended item output IF 313 receives a list of recommended items and its score from the item score calculation means 325. The recommended item output IF 313 transmits the list of recommended items and the score thereof to the user terminal 21 of the user 1 who transmitted the item recommendation request. The user 1 operates the user terminal 21 to display a list of recommended items on a display screen or the like.

It is assumed that the user 1 operates the user terminal 21 and refers to a favorite item i from the recommendation list. The reference to the item i may be performed by an action execution unit provided outside, or may be performed by an action execution unit in the recommendation engine 30. The user terminal 21 notifies the user action input IF 311 that the user 1 referred to the item i. The user action input IF 311 changes the corresponding element (i-th element) of the action history X ₁ of the user 1 in the user action history DB 321 to 1 and transmits it to the weight calculation means 322.

  The weight calculation unit 322 calculates the weight with reference to the user action history DB 321. Further, the correlation calculation unit 326 refers to the user action history DB 321 and calculates a correlation coefficient between the user pairs. The user pair relationship calculation unit 323 calculates the product of the weight calculated by the weight calculation unit 322 and the correlation coefficient calculated by the correlation calculation unit 326, and sets this as the relationship between the user pairs. As the user 1 action history stored in the user action history DB 321 is changed, the weight calculated by the weight calculating unit 322 and the correlation coefficient calculated by the correlation calculating unit 326 change and are the product of them. The relationship between user pairs changes. The recalculation of weights and relationships may be performed every time the user's action history changes, or may be performed when a certain amount of change has been accumulated. Or you may perform regularly.

  In the present embodiment, the user pair relationship calculating unit 323 refers to the user behavior history DB 321 and calculates a relationship between pairs in which asymmetry is allowed between user pairs. In this embodiment, since the item score is calculated using the relationship in which asymmetry is allowed between user pairs, collaborative filtering that more accurately reflects the relationship between user pairs is possible.

  In this embodiment, the weight calculation unit 322 calculates a weight having asymmetry between user pairs, and the correlation calculation unit 326 calculates a correlation coefficient having symmetry between user pairs. By making the product of the weight and the correlation coefficient the relationship between the user pairs, the relationship between the user pairs can be made asymmetric.

  When weights are viewed from one side of a user pair and the other user's reference item is considered to be helpful, the weight to the other user as viewed from one user is increased, and vice versa. When it is considered that the reference item of one user is not helpful, the weight to one user viewed from the other user may be reduced. In this embodiment, when viewed from a certain user, the weight is reduced when there are many items that only the other user of the pair refers to, and the weight is increased when the number of items referenced only by the other user is small. By performing such weight setting, it is possible to suppress the degree of influence of a user who has a strong relationship with anyone, such as a user who has an action history in various items.

  For example, considering the above-described pair of user 1 and user 2, the number of action histories of user 1 is 110, and the number of action histories of user 2 is 10010. The user 1 is considered to be a user who refers to an item in a specific field, and the user 2 is considered to be a user who refers to various items. When the asymmetry of the relationship is not taken into account, the item score of an item having an action history only for the user 1 and the item score of an item having an action history only for the user 2 are approximately the same. Since the preference (reference range) of the user 1 has a large overlap with the preference of the user 2, an item having an action history only in the user 1 should be recommended to the user 2. However, the user 2 has many action histories with different preferences from the user 1, and an item having an action history only for the user 2 is not so meaningful to the user 1 even if it is recommended. In the present embodiment, the relationship between the user pairs is asymmetrical, the relationship of the user 2 viewed from the user 1 is weakened, and the relationship of the user 1 viewed from the user 2 can be strengthened. Therefore, it is possible to recommend an item having an action history only to the user 1 to the user 2 and not to recommend an item having an action history only to the user 2 to the user 1.

と定義し、重み算出手段３２２は、

を用いて重みを算出してもよい。例えば、ユーザｐとユーザｑとが共通に参照したアイテムが１００あるとして、そのうち、ｐが先に参照したアイテム数が３０で、ｑが先に参照したアイテム数が７０であれば、
ｗ_ｐｑ＝８５、ｗ_ｑｐ＝６５
となる。このような重みを用いる場合、ユーザペア間で共通に参照したアイテムについて、より早くアイテムを参照したユーザが重要視され、影響力を持つことになる。 In the above, the weight is set to the weight according to the number of items having an action history only on the other side of the user pair, but the weight is not limited to this. For example, a method of increasing the weight of the earlier reference to the item may be used. Specifically, for each item, a _pq (i)

The weight calculation means 322 is defined as

The weight may be calculated using For example, if there are 100 items commonly referred to by the user p and the user q, if the number of items referred to by p is 30 and the number of items referred to by q is 70,
w _pq = 85, w _qp = 65
It becomes. When such a weight is used, the user who referred to the item earlier is emphasized and has an influence on the item commonly referred to between the user pairs.

  In the first embodiment, collaborative filtering is U2UCF, but it can also be applied to other relationships such as I2ICF. For example, in the case of I2ICF, the weight between item pairs is calculated in the same manner as the weight between user pairs, and this is used to allow a one-to-one asymmetry between item pairs. May be calculated. Alternatively, the user's presence (information such as location and position) may be used to calculate the weight between the context pairs, and this may be used to calculate the relationship between the context pairs. For example, for user 1, the presence at work is context 1 and the private presence is context 2. Further, for the user 2, the presence of being at work is defined as context 3, and the private presence is defined as context 4. When using context, for example, context 1 (working user 1) has a great influence on context 2 (private user 1), but the reverse is not true, and collaborative filtering using relationships is possible. become.

  Next, the second embodiment will be described. The configuration of the collaborative filtering system of the second embodiment of the present invention is the same as that of the collaborative filtering system of the first embodiment shown in FIG. In the collaborative filtering system of the present embodiment, the correlation calculation unit 326 sets the influence of the element on the calculation of the correlation coefficient between the pairs according to the element with respect to the element that is the calculation source of the correlation coefficient between the pairs. This is different from the first embodiment in that the correlation coefficient is calculated above. That is, in the first embodiment, the correlation calculation unit 326 calculates the correlation coefficient using Equation 2, and the influence of individual items on the correlation coefficient is constant between items. In the present embodiment, it is possible to set the influence of the item on the calculation result of the relationship according to the item.

式５におけるｋ（ｉ）は、ｉ番目のアイテムが関係性算出に与える影響を表す値（係数）である。ｋ（ｉ）は、例えば、０≦ｋ（ｉ）≦１の値を採る。ｋ（ｉ）の値が大きいほど、ｉ番目のアイテムが相関係数に与える影響は大きくなり、ｋ（ｉ）＝０のとき、ｉ番目のアイテムは、相関係数の計算に影響しない。全てのｋ（ｉ）がｋ（ｉ）＝１のとき、式５を用いて計算した相関係数は、Ｐｅａｒｓｏｎ相関係数と等しくなる。なお、上記では、Ｐｅａｒｓｏｎ相関係数の算出式にｋ（ｉ）を導入しているが、ベースとなる相関係数の算出式は任意であり、Ｐｅａｒｓｏｎ相関係数の算出式には限定されない。 The correlation calculation means 326 calculates the correlation coefficient of the user pair (user p, user q) using the following formula 5.

K (i) in Equation 5 is a value (coefficient) representing the influence of the i-th item on the relationship calculation. For example, k (i) takes a value of 0 ≦ k (i) ≦ 1. The larger the value of k (i), the greater the influence of the i-th item on the correlation coefficient. When k (i) = 0, the i-th item does not affect the calculation of the correlation coefficient. When all k (i) are k (i) = 1, the correlation coefficient calculated using Equation 5 is equal to the Pearson correlation coefficient. In the above description, k (i) is introduced into the Pearson correlation coefficient calculation formula, but the base correlation coefficient calculation formula is arbitrary, and is not limited to the Pearson correlation coefficient calculation formula.

  The value of k (i) is determined as follows, for example. Information such as a category to which the item belongs is stored as meta information in the item data storage DB 34. Correlation calculation means 326 determines the value of k (i) according to the category of the item. For example, when transmitting a recommendation request for an item to the recommendation request receiving unit 312, the user designates an item category that is more important. Correlation calculation means 326 sets k (i) to 1 for items in categories that the user places importance on, and halves the value of k (i) for items in similar categories. K (i) is set to 0. The value of k (i) described here is merely an example, and the method of setting the value of k (i) for the i-th item is arbitrary.

The calculation of the weight w _pq performed by the weight calculation unit 322 is the same as in the first embodiment. That is, the weight calculation unit 322 calculates a weight that allows asymmetry. The relationship calculating unit 323 calculates the relationship between the user pairs by the following equation 6 using the weight calculated by the weight calculating unit 322 and the correlation coefficient calculated by the correlation calculating unit 326.

In the present embodiment, the correlation calculation unit 326 sets the influence of the element on the calculation of the correlation coefficient between the pairs for the element that is the calculation source of the correlation coefficient between the pairs, Calculate the number of relationships. By doing in this way, about the specific item, the influence which the action information with respect to the item has on the calculation of the relationship between the user pairs can be increased. The influence on the calculation of the relationship between the pairs can be reduced. Also in the present embodiment, the weight calculation means 322 calculates the relationship C (X _p , X _q ) with the user q as viewed from the user p using w _pq , as in the first embodiment. It is possible to calculate the relationship reflecting the asymmetry between users.

  In the second embodiment, the collaborative filtering is U2UCF. However, as in the first embodiment, it can be applied to other relationships such as I2ICF. For example, in the case of I2ICF, the correlation calculation unit 326 sets the influence of the user on the calculation of the correlation coefficient between item pairs according to the user for the user who is the element from which the correlation coefficient is calculated. What is necessary is just to calculate a correlation coefficient. Also, when calculating the relationship between context pairs, the correlation calculation means 326 uses the item as an element that is the element from which the correlation coefficient is calculated. The correlation coefficient may be calculated after setting it accordingly. The calculation of the weight is as described in the first embodiment.

  A third embodiment of the present invention will be described. The configuration of the collaborative filtering system of the third embodiment of the present invention is the same as that of the collaborative filtering system of the first embodiment shown in FIG. In the second embodiment, the correlation calculation unit 326 can set the influence of the element on the calculation of the correlation coefficient between the pairs according to the item. This embodiment is different from the second embodiment in that the magnitude of the influence of elements on the calculation of the correlation coefficient between pairs allows asymmetry between the pairs. In the present embodiment, asymmetry is allowed not only in the weight calculated by the weight calculation unit 322 but also in the correlation coefficient calculated by the correlation calculation unit 326.

式７におけるｋ_ｐｑ（ｉ）は、ｉ番目のアイテムが関係性算出に与える影響を表す値（係数）である。ｉ番目のアイテムについて、ユーザｐから見たユーザｑの相関係数を算出するときと、ユーザｑから見たユーザｐの相関係数を算出するときとで、ｉ番目のアイテムが相関係数の算出に与える影響は相互に異なることが許容される。すなわち、ｋ_ｐｑ（ｉ）≠ｋ_ｑｐ（ｉ）となることが許容される。ｋ_ｐｑ（ｉ）は、例えば、０≦ｋ_ｐｑ（ｉ）≦１の値を採る。ｋ_ｐｑ（ｉ）の値が大きいほど、ｉ番目のアイテムが相関係数に与える影響は大きくなり、ｋ_ｐｑ（ｉ）＝０のとき、ｉ番目のアイテムは、相関係数の計算に影響しない。全てのｋ_ｐｑ（ｉ）がｋ_ｐｑ（ｉ）＝１のとき、式７を用いて計算した相関係数は、Ｐｅａｒｓｏｎ相関係数と等しくなる。また、全てのアイテムについて、ｋ_ｐｑ（ｉ）＝ｋ_ｑｐ（ｉ）のとき、ユーザｐから見たユーザｑの相関係数と、ユーザｑから見たユーザｐの相関係数とは等しくなる。 The correlation calculation unit 326 calculates the correlation coefficient of the user q viewed from the user p using the following equation (7).

K _pq (i) in Expression 7 is a value (coefficient) representing the influence of the i-th item on the relationship calculation. For the i-th item, when the correlation coefficient of the user q viewed from the user p is calculated and when the correlation coefficient of the user p viewed from the user q is calculated, the i-th item is the correlation coefficient The influence on the calculation is allowed to be different from each other. That is, k _pq (i) ≠ k _qp (i) is allowed. k _pq (i) is, for example, takes a value of _{0 ≦ k pq (i) ≦} 1. The larger the value of k _pq (i), the greater the influence of the i-th item on the correlation coefficient. When k _pq (i) = 0, the i-th item does not affect the calculation of the correlation coefficient. . When all k _pq (i) are k _pq (i) = 1, the correlation coefficient calculated using Equation 7 is equal to the Pearson correlation coefficient. For all items, when k _pq (i) = k _qp (i), the correlation coefficient of the user q viewed from the user p is equal to the correlation coefficient of the user p viewed from the user q.

The value of k _pq (i) is determined as follows, for example. When the user q has a reference history for the i-th item, k _pq (i) = 1, otherwise k _pq (i) = 1/2. In this case, the influence that the item having the reference history of the user q has on the calculation of the correlation coefficient is increased. Alternatively, when the user q refers to the i-th item before the user p, k _pq (i) = 1, otherwise k _pq (i) = 1/2. In this case, it is possible to increase the influence of the item referred to by the user q before the user p on the calculation of the correlation coefficient of the user q viewed from the user p. The value of k _pq (i) described here is merely an example, and the method of setting the value of k _pq (i) for the i-th item is arbitrary. When calculating the correlation coefficient, filtering may be performed in advance, and the value of k _pq (i) may be set to 0 for items belonging to a specific category.

The calculation of the weight w _pq performed by the weight calculation unit 322 is the same as in the first embodiment. That is, the weight calculation unit 322 calculates a weight that allows asymmetry. The relationship calculating unit 323 calculates the relationship between the user pairs by the following equation 8 using the weight calculated by the weight calculating unit 322 and the correlation coefficient calculated by the correlation calculating unit 326.

In the present embodiment, when calculating the correlation coefficient, the magnitude of the influence of the element on the calculation of the correlation coefficient between the pairs allows asymmetry between the pairs. By doing in this way, the influence which the action information with respect to an item has on the calculation of the relationship between user pairs can be changed when the other is viewed from one of the pairs and when the other is viewed from the other. Also in the present embodiment, the relationship C (X _p , X _q ) with the user q as viewed from the user p is calculated using w _pq calculated by the weight calculation unit 322, and similarly to the first embodiment. It is possible to calculate the relationship reflecting the asymmetry between users.

  In the third embodiment, U2UCF is used for collaborative filtering, but it can also be applied to other relationships such as I2ICF as in the second embodiment. For example, in the case of I2ICF, the correlation calculation unit 326 sets the influence of the user on the calculation of the correlation coefficient between the item pairs asymmetrically with respect to the user who is the element from which the correlation coefficient is calculated. May be calculated. When calculating the relationship between the context pairs, the correlation calculating unit 326 makes the influence of the item on the calculation of the correlation coefficient between the context pairs asymmetrical with respect to the item that is the element from which the correlation coefficient is calculated. Thus, the correlation coefficient may be calculated. The calculation of the weight is as described in the first embodiment.

By the way, in the first to third embodiments, if the weight between the pair for calculating the relationship and the influence of the item on the correlation calculation are all set to 1, the relationship between the pair becomes symmetric. That is, if the weights w _pq in equation 2 are all 1, the weights w _pq and k (i) in equation 6 are all 1, and w _pq and k _pq (i) in equation 8 are all set to 1, The relationship is symmetric. This means that it is equivalent to existing collaborative filtering. Therefore, if the operations of the weight calculation means and the correlation calculation means are changed, it is possible to switch between collaborative filtering using a relationship having asymmetry and collaborative filtering using a relationship having symmetry.

  As mentioned above, although this invention was demonstrated based on the suitable embodiment, the collaborative filtering system of this invention, a method, and a program are not limited only to the said embodiment, Various from the structure of the said embodiment. Those modified and changed as described above are also included in the scope of the present invention.

  Finally, the minimum configuration of the present invention will be described. The collaborative filtering system of the present invention has a relationship calculation means and an item score calculation means as a minimum configuration. The relationship calculation means refers to the user action history DB that stores history information of actions performed by the user on the item, and a one-to-one asymmetry is allowed, between user pairs, between item pairs, or context Calculate the relationship between pairs. The item score calculation means calculates the item score using the relationship between the pair and the action history information.

  The minimum configuration of the collaborative filtering method of the present invention refers to a user action history DB that stores history information of actions performed by a user on an item, and a one-to-one asymmetry is allowed. Or a step of calculating a relationship between context pairs, and a step of calculating an item score using the relationship between the pairs and action history information.

  The minimum configuration of the program of the present invention refers to a user action history DB that stores history information of actions performed by a user on an item in a computer, and a one-to-one asymmetry is allowed. In this configuration, a process for calculating the relationship between pairs or context pairs and a process for calculating an item score using the relationship between the pairs and the action history information are executed.

  In the present invention, a relationship in which asymmetry is allowed is calculated, and an item score is calculated using the relationship. As for the relationship between users, one-to-one asymmetry is allowed, and between the pair, the relationship when looking from one to the other is strengthened, and the relationship when looking from the other to the other is weakened. Is possible. Therefore, by using such a relationship for collaborative filtering, collaborative filtering that considers asymmetry of the strength of the relationship between the calculated pairs becomes possible.

21 to 23: User terminal 30: Recommendation engine 31: User IF
32: Scoring unit 34: Item data storage DB
35: User data storage DB
311: User action input IF
312: Recommendation request reception IF
313: Recommended item output IF
321: User action history DB
322: Weight calculation unit 323: User pair relationship calculation unit 324: User pair relationship storage DB
325: Item score calculating means 326: Correlation calculating means

Claims (18)

Refer to the user action history DB that stores history information of actions performed by the user on the item, and the relationship between user pairs, between item pairs, or between context pairs, where one-to-one asymmetry is allowed A relationship calculating means for calculating;
A collaborative filtering system comprising an item score calculating means for calculating an item score using the relationship between the pair and the action history information.   The relationship calculation means refers to the user action history DB, refers to the weight calculation means for determining a weight that allows asymmetry for a pair for calculating the relationship, and refers to the user action history DB, Correlation calculating means for calculating a correlation coefficient indicating the correlation between the pairs, and the product of the weight determined by the weight calculating means and the correlation coefficient calculated by the correlation calculating means is the relationship between the pairs The collaborative filtering system according to claim 1.   The correlation calculation means calculates a correlation coefficient after setting the influence of the element on the calculation of the correlation coefficient between the pairs according to the element with respect to the element from which the correlation coefficient between the pairs is calculated. The collaborative filtering system according to claim 2.   The collaborative filtering system according to claim 3, wherein the magnitude of the influence of the element on the calculation of the correlation coefficient between the pairs allows an asymmetry between the pairs.   The weight calculating means, when viewed from one of the pairs, between the pairs for calculating the relationship, the weight to the other viewed from one of the pairs based on the amount of behavior history existing only in the other of the pair The collaborative filtering system according to any one of claims 2 to 4, wherein:   The collaborative filtering system according to any one of claims 2 to 4, wherein the weight calculation unit determines the weight based on a time history relationship between pairs for calculating the relationship. A computer refers to a user behavior history DB that stores history information of actions performed on an item, and a one-to-one asymmetry is allowed, between user pairs, between item pairs, or between context pairs Calculating a relationship;
A collaborative filtering method comprising: a step of calculating an item score using the relationship between the pair and the action history information. Calculating the relationship comprises:
A weight calculating step in which the computer refers to the user behavior history DB and determines a weight at which asymmetry is allowed for a pair for calculating a relationship;
A correlation calculation step in which the computer refers to the user behavior history DB and calculates a correlation coefficient indicating a correlation between the pairs;
The collaborative filtering method according to claim 7, further comprising: calculating the product of the weight and the correlation coefficient to obtain a relationship between the pairs.   In the correlation calculation step, the computer sets the influence of the element on the calculation of the correlation coefficient between the pairs according to the element with respect to the element from which the correlation coefficient between the pairs is calculated. The collaborative filtering method according to claim 8, wherein the number is calculated.   The collaborative filtering method according to claim 9, wherein the magnitude of the influence of the element on the calculation of the correlation coefficient between the pairs allows an asymmetry between the pairs.   In the weight calculation step, the computer is viewed from one of the pairs based on the amount of behavior history existing only in the other of the pair, when viewed from one of the pair between the pairs for calculating the relationship. The collaborative filtering method according to claim 8, wherein a weight for the other is determined.   The cooperation according to any one of claims 8 to 10, wherein in the weight calculation step, the computer determines the weight based on a time history relationship between pairs of behavior history between the pairs for calculating the relationship. Filtering method. On the computer,
Refer to the user action history DB that stores history information of actions performed by the user on the item, and the relationship between user pairs, between item pairs, or between context pairs, where one-to-one asymmetry is allowed Processing to calculate,
A program for executing a process of calculating an item score using the relationship between the pair and the action history information. The process of calculating the relationship
A weight calculation process that refers to the user behavior history DB and determines a weight at which asymmetry is allowed for a pair for calculating a relationship;
A correlation calculation process for calculating a correlation coefficient indicating a correlation between the pairs with reference to the user behavior history DB;
The program according to claim 13, further comprising: calculating a product of the weight and the correlation coefficient to obtain a relationship between the pairs.   In the correlation calculation process, the correlation coefficient is calculated after setting the influence of the element on the calculation of the correlation coefficient between the pairs for the element that is the calculation source of the correlation coefficient between the pairs. The program according to claim 14.   The program according to claim 15, wherein the magnitude of the influence of the element on the calculation of the correlation coefficient between the pairs allows an asymmetry between the pairs.   In the weight calculation process, when viewed from one of the pairs between the pairs for calculating the relationship, the weight to the other viewed from one of the pairs based on the amount of behavior history existing only in the other of the pair The program according to claim 14, wherein the program is determined.   The program according to any one of claims 14 to 16, wherein, in the process of calculating the weight, the weight is determined based on a time history of an action history between a pair for calculating the relationship.

Images