JP2013257793A - Commodity diversification recommendation device, and method and program for the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a recommendation list with an extremely high level of satisfaction of a user.SOLUTION: In the present invention, a recommendation term which is a probability of purchase of each commodity by each user is calculated by using a purchase history of each user ID. From the purchase history, a diversification term which is a ratio of users who have purchased a certain commodity to users who have not purchased a certain commodity group is calculated, and a recommendation history showing that which commodity group is provided as a recommendation list to which user, that which commodity the user has browsed its detail information in the recommendation list or the user has not browsed detail information on any commodities, is obtained. By using the recommendation term, the diversification term, and the recommendation history, a diversification parameter which is a weight of a diversification term with respect to a recommendation term at the time of calculating a recommendation score when a recommendation list is created is calculated, and the recommendation list is created by using the purchase history, the recommendation term, the diversification term, the recommendation history, and the diversification parameter.

Description

  The present invention relates to a product recommendation device, method, and program, and more particularly, to a product recommendation device, method, and program in user product purchase provided by an E-Commerce (electronic commerce) service.

  A number of information recommendation methods have been proposed. For example, in order to recommend a news article, the user is given a five-step evaluation for each news article, the past evaluation history is viewed, the degree of preference similarity between users is measured, and the news article preferred by a user with a similar preference Has been proposed (see, for example, Non-Patent Document 1).

  Also, let the user specify the text of the part of interest, estimate the user's interest based on the appearance frequency of the words contained therein, search the information in the community based on the word representing the user's interest, There is a method of presenting only information of a portion that is of interest to the user (for example, see Patent Document 1).

  In order to maximize user satisfaction, diversification of recommendation lists is effective. For example, when recommending movies, it is known that even if users are sci-fi enthusiasts, it is more satisfying to recommend sci-fi, horror, and action than to recommend only sci-fi. (For example, refer nonpatent literature 2).

JP 11-338869 A

Resnick, P. et al., "GroupLens: An open architecture for collaborative filtering of netnews," CSCW '94: Proceedings of the 1994 ACM conference on Computer supported cooperative work, ACM Press New York, 1994. Ziegler C.N. et al, `` Improving recommendation lists through topic diversification, '' Proceedings of the 14th International Conference on World Wide Web, ACM Press New York, 2005

  In general, the products handled by the E-Commerce service are wide, and it is difficult for users to find products that they are really interested in.

  When a product is recommended using the method of Patent Literature 1 described above, a product related to an interesting word designated by the user is recommended. In this case, there is a problem that a word must be specified. Such a system that requires user interaction is inconvenient for the user.

  When a product is recommended by the method described in Non-Patent Document 1, a user's preference is estimated and a product associated with it is recommended, so that only similar products are recommended and the user's satisfaction is lowered. In addition, since it is highly likely that the recommended product is known to the user, it is difficult for the recommended product to lead to a new purchase.

  In addition, if the product is recommended by the method described in Non-Patent Document 2, the user's satisfaction is improved, but the parameter of how diversified the recommendation list needs to be manually set. It is difficult to find a parameter that maximizes.

  The present invention has been made in view of the above points. In the E-Commerce service, the user satisfaction is determined by using information on which product the user has selected in the recommendation list or which product has not been recommended. It is an object of the present invention to provide a product diversification recommendation apparatus, method, and program capable of providing a recommendation list with a very high degree of user satisfaction by estimating a recommendation list diversification parameter that maximizes the degree.

In order to solve the above problem, the present invention (Claim 1) is a product diversification recommendation device that uses a user's purchase history and recommendation history in an electronic commerce service (E-Commerce) service,
A purchase history acquisition unit that acquires a purchase history of each user ID in the E-Commerce service and stores it in a purchase history storage unit;
Using the acquired purchase history, a recommended item that is a probability of purchasing each product of each user is calculated, and a recommended item calculating unit that is stored in a recommended item storage unit;
Diversification term calculation means for calculating a diversification term that is a ratio of users who have purchased a certain product among users who have not purchased a certain product set using the acquired purchase history;
In the E-Commerce service, a recommendation regarding to which user, what product set is a recommendation list, and which detailed information of which product in the recommendation list has been viewed, or which detailed information of any product has not been browsed A recommendation history acquisition means for acquiring a history and storing the history in a recommendation history storage means;
Using the calculated recommendation term, the calculated diversification term, and the acquired recommendation history, a diversification parameter that is a weight of the diversification term with respect to the recommendation term at the time of recommendation score calculation when creating a recommendation list is calculated Diversification parameter calculation means for storing in the diversification parameter storage means;
A recommendation list creating means for creating a recommendation list by using the acquired purchase history, the calculated recommendation item, the calculated diversification item, the acquired recommendation history, and the calculated diversification parameter; Have.

Further, the present invention (Claim 2) provides the diversification term calculation means,
Using the purchase history acquired by the purchase history acquisition means, calculate a popularity term that is a probability that each product is purchased, and store it in a popularity term storage means;
Using the acquired purchase history, for each pair of two products, a dissimilarity term is calculated, which is the proportion of users who have purchased one product but who have not purchased the other product. And a dissimilarity term calculation means for storing in the dissimilarity term storage means,
Using the acquired purchase history, for each pair of two products, a similarity term that is the proportion of users who have not purchased one product among the users who have not purchased one product is calculated. , Similarity term calculation means stored in the similarity term storage means,
Means for calculating a diversification term using the calculated popularity term, the calculated dissimilarity term, and the calculated similarity term.

  The present invention provides a recommendation history (which user is logged in (or is doing), what product is recommended, which product is clicked, or which product is not clicked. Data) is used to optimize the diversification parameter, the user satisfaction, that is, the probability of purchasing any one of the products in the recommendation list is higher than that of the conventional recommendation technique.

It is an example of a recommendation list posting page. It is a block diagram of the merchandise information recommendation apparatus in one embodiment of this invention. It is an example of the purchase history table in one embodiment of the present invention. It is an example of the recommendation term table in one embodiment of this invention. It is an example of the popularity degree term table in one embodiment of this invention. It is an example of the dissimilarity term table in one embodiment of the present invention. It is an example of the similarity term table in one embodiment of this invention. It is an example of the recommendation history table in one embodiment of the present invention. It is a flowchart at the time of the purchase history process in one embodiment of this invention. It is a flowchart at the time of the recommendation term calculation in one embodiment of this invention. It is a flowchart of the detailed process of step 230 of FIG. 10 in one embodiment of this invention. It is a flowchart of the detailed process of step 232 of FIG. 11 in one embodiment of this invention. It is a flowchart at the time of the diversification term calculation in one embodiment of this invention. It is a flowchart of the detailed process of step 320 of FIG. 13 in one embodiment of this invention. It is a flowchart of the detailed process of step 330 of FIG. 13 in one embodiment of this invention. It is a flowchart of the detailed process of step 340 of FIG. 13 in one embodiment of this invention. It is a flowchart at the time of the recommendation log | history process in one embodiment of this invention. It is a flowchart at the time of the diversification parameter estimation in one embodiment of this invention. It is a flowchart at the time of the recommendation list preparation in one embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  First, the outline of the present invention will be described.

  An object of the present invention is to provide a diversified recommendation list that maximizes user satisfaction. Here, the satisfaction degree of the user is defined as a probability that “when a recommendation list is presented, at least one of them is purchased”. That is, instead of simply arranging products in descending order of probability of purchase by the user as in the prior art, if the length of the recommended list to be provided is L, any one of the L products Create a recommendation list to maximize your purchase probability.

  The first one is the product with the highest probability of purchase by the user. The second-ranked product is not simply the product with the second highest probability of purchase by the user, but the probability of purchasing using the information that the first-ranked product is not purchased or not liked at the time of selecting the second-ranked item. calculate. In this way, the second-ranked product becomes a product most likely to be purchased by the user in a genre different from the first-ranked product, and when applied recursively, the recommendation list is diversified.

  The explanation will be made more strictly using mathematical expressions. L items for user u

を推薦するものとする。1位には、従来の協調フィルタリングと変わらず、

Shall be recommended. The first place is the same as conventional collaborative filtering,

ここでy_lは推薦リスト中のｌ番目の商品が購買されるかどうかを表す2値の変数であり、y_l=1は1番目の商品が購買される場合を、y_l=0は購買されない場合を表す。すなわち該ユーザの購買する確率の最も高い商品を配置する。２位には、従来の協調フィルタリングならば該ユーザの購買する確率が1番目に高いものを配置するが、本発明では、

Here, y _l is a binary variable indicating whether or not the l th product in the recommendation list is purchased, y _l = 1 is the case where the first product is purchased, and y _l = 0 is the purchase. Indicates the case where it is not done. That is, the product with the highest probability of purchase by the user is placed. In the second place, in the case of conventional collaborative filtering, the one having the highest probability of purchase by the user is placed. In the present invention,

となる商品を推薦する。すなわち、1位の商品は購買しないという情報を用いる。本情報を考慮すれば、2位に配置する商品は1位に配置された商品とは異なるジャンルのものになりやすくなる。ｌ位も再帰的に以下のように決定される。

Recommending a product. That is, information that the first-ranked product is not purchased is used. Considering this information, the product placed in the second place tends to be of a different genre from the product placed in the first place. The 1st place is also recursively determined as follows.

ここで

here

ただし

However,

は空集合、

Is an empty set,

ただし、

However,

は空集合である。

Is an empty set.

Next, a method for calculating the right side of Equation (3) will be described. Probability that user u purchases product i _l because it is difficult to find the right side directly

とｌ−1位までを購買しない場合に商品ｉ_lを購買する確率

And the probability of purchasing the product i ₁

から求める。前者の確率は通常の推薦スコアであるため、推薦項と呼ぶことにし、後者の確率は推薦リストの多様化に影響を与えるため多様化項と呼ぶことにする。

Ask from. Since the former probability is a normal recommendation score, it will be called a recommendation term, and the latter probability will be called a diversification term because it affects the diversification of the recommendation list.

  Now, the right side of Equation (3) is defined as follows using the recommendation term and the diversification term.

ここでαが多様化パラメータであり、このパラメータの値が大きいほど多様化されることになる。作成した推薦リストのうち１つでも購買される確率が最大になるように、αを最適化するのが本発明のポイントである。

Here, α is a diversification parameter, and the greater the value of this parameter, the greater the diversification. The point of the present invention is to optimize α so that the probability of purchasing one of the prepared recommendation lists is maximized.

  It is assumed that a recommendation system is operated on the E-Commerce service. The user logs in and visits a page where the recommendation list as shown in FIG. 1 can be viewed. Each product in the recommendation list is given a link to a page of detailed information on the product. Each time a user moves from the recommendation list page to another page, which user was logged in (or is doing), what product was recommended, which product was clicked, and which It is assumed that history data (hereinafter referred to as recommendation history) relating to whether a product has not been clicked is accumulated.

Α is determined so that the logarithmic likelihood (which is the probability that exactly the same data is generated) of which product in the accumulated recommendation history is clicked or not is maximized. Assume that R recommendation histories are accumulated. In the r = (1,..., R) recommendation history, the recommended user is u _r , the recommended product is i _{r, l} (l = 1,..., L), and the clicked position is l _r ε ( [1, L], 0) if none is clicked. At this time, the generation probability of which item in the r-th recommendation history has been clicked or not can be expressed as in Equation (5).

このとき、推薦履歴の対数尤度は下式のように表すことができる。

At this time, the log likelihood of the recommendation history can be expressed as the following equation.

前記対数尤度を最大にするは、最急降下法等の最適化手法によって求める。

The log likelihood is maximized by an optimization method such as the steepest descent method.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 2 shows a configuration of a product recommendation device according to an embodiment of the present invention.

  The product recommendation device 1 shown in FIG. 1 includes a purchase history processing unit 10, a recommendation item calculation unit 20, a diversification item calculation unit 30, a recommendation history processing unit 40, a diversification parameter estimation unit 50, a recommendation list creation unit 60, and a recording unit. 70, a communication unit 80, and an input / output unit 90. The communication unit 80 is connected to the network 3, and the input / output unit 90 is connected to the input device or the external device 4 of the display device. The external terminal 2 corresponds to an EC site that cooperates with the present system, for example.

  The storage unit 70 includes a purchase history table 71, a recommended item table 72, a diversification item recording unit 73, and a diversification parameter recording unit 74. The diversification term recording unit 73 includes a popularity term table 731, a dissimilarity term table 732, and a similarity term table 733. Each table will be described below.

<Purchase history table 71>
As shown in FIG. 3, the purchase information table 71 includes a user ID field and a product ID field.

  In the user ID field, an identifier for identifying a user who has purchased the product is set by the purchase history processing unit 10.

  The product ID field corresponds to an identifier of a product that appears in a product information table (not shown), and is set by the purchase history processing unit 10.

<Recommended Item Table 72>
As shown in FIG. 4, the recommended item table 72 includes a user ID field, a product ID field, and a recommendation score field.

  In the recommendation score field, a recommendation term (probability that the user u purchases the product i) P (y = 1 | u, i) of the user u is set by the recommendation term calculation unit 20.

<Popularity item table 731>
As shown in FIG. 5, the popularity degree table 731 includes a product ID field and a popularity degree field.

  In the popularity degree field, the diversification term calculation unit 30 sets the popularity degree (probability that the commodity i is purchased) P (y = 1 | i) of the commodity i.

<Dissimilarity term table 732>
As shown in FIG. 6, the dissimilarity term table 732 includes a product ID field, an upper product ID field, and a dissimilarity field.

  Here, the dissimilarity literally represents the dissimilarity between the two products specified by the product ID and the higher-level product ID. However, if the dissimilarity is strictly defined, the product specified in the product ID field is determined by the user who buys the product. Among them, the product specified in the upper product ID field indicates the percentage of users who do not buy.

  In the upper product ID field, it corresponds to the identifier of the product that appears in the purchase history table, and is set by the diversification item calculation unit 30.

  The dissimilarity field is a dissimilarity between the product i and the top product i ′ (probability of not purchasing the top product i ′ among users purchasing the product i) P (y ′ = 0 | i ′, y = 1, i) is set by the diversification term calculation unit 30.

<Similarity term table 733>
As shown in FIG. 7, the similarity term table 733 includes a product ID field, an upper product ID field, and a similarity field.

  Here, the similarity is literally the similarity between two products specified by the product ID and the higher-level product ID. However, when the similarity is strictly defined, among users who do not purchase the product specified in the product ID field. This also indicates the percentage of users who do not purchase the product specified in the upper product ID field.

  The similarity field indicates the similarity between the product i and the higher-order product i ′ (the ratio of users who do not purchase the product i among the users who do not purchase the higher-order product i ′) P (y ′ = 0 | i ′, y = 0, i) is set by the diversification term calculation unit 30.

<Recommendation history table 74>
As shown in FIG. 8, the recommendation history table 74 includes a user ID field, a click position field, and a product ID field 1... Here, L represents the length of the recommendation list, and the value of L is set in advance by the operator. FIG. 8 shows a case where L = 3. Note that the number of product ID fields depends on the length of the recommendation list.

  The user ID field corresponds to the user ID appearing in the purchase history table, and to which user the recommendation list was created is set by the recommendation history processing unit 40.

  The click position field indicates which position of the link to the detailed information page of the product in the L recommendation lists specified in the product ID field 1... L is clicked (but none of the links are clicked). In this case, 0 is set), and is set by the recommendation history processing unit 40.

  The product ID field 1... L indicates which product is recommended among the product IDs appearing in the purchase history table, and is set by the recommendation history processing unit 40.

  First, the processing at the time of purchase history processing in one embodiment of the present invention will be described in detail. In the present embodiment, the purchase history is transmitted to the product recommendation device 1 via the network 3 and the communication unit 90 from the EC site linked each time the user purchases the product, and this processing is executed.

  FIG. 9 is a flowchart of processing during purchase history processing according to an embodiment of the present invention.

  Step 110) The purchase history processing unit 10 inserts in the purchase history table 71 a row in which values of the user ID field and the product ID field are set according to the purchased user and the purchased product. It is assumed that this process occurs every time a user purchases some product.

  Next, processing at the time of calculating a recommended term in an embodiment of the present invention will be described. In the present embodiment, this processing is performed when the system administrator uses the external device 4 to issue a recommendation term calculation command to the product recommendation device 1. It may be possible to set it to be executed automatically.

  FIG. 10 is a flowchart of a process at the time of calculating a recommended item according to an embodiment of the present invention.

  Step 210) The recommended term calculation unit 20 sets parameters Z, α, and β for calculating the recommended term. These parameters are set by constants given in advance. However, Z is an integer of 1 or more, and α and β are decimal numbers of 0 or more.

  Step 220) The recommended item calculation unit 20 refers to the purchase information table 51, and the user of each purchase

と、商品

And product

を取得する。ここでMは購買履歴の総数であり、mは各購買履歴を一意に表すIDである。

To get. Here, M is the total number of purchase histories, and m is an ID that uniquely represents each purchase history.

  Step 230) The recommended term calculation unit 20 uses the Collapsed Gibbs sampling described in the document “Bishop, M.C.,“ Pattern recognition and machine learning ”, 2006” to identify each purchase topic.

を推定する。

Is estimated.

  Step 240) Using the parameters Z, α, and β set in Step 210 by the recommended term calculation unit 20 and the topic z of each purchase estimated in Step 230, each user u The probability P (i | u) for selecting each product i is calculated.

ここでM_zはトピックzにアサインされた購買数、M_ziはトピックzにアサインされ、かつ商品iに関する購買の数、M_uzはたユーザuの購買の中でトピックzにアサインされたものの数、M_uはユーザuの購買数を表す。M_zとM_zi、M_uzはステップ２３０にて推定した各購買のトピックｚから求めることができる。なお、Ｕはユーザの総数を表す。

Where M _z is the number of purchases assigned to topic z, M _zi is the number of purchases for product i and M _{uz is} the number of purchases of user u assigned to topic z , M _u represents the number of purchases of the user u. M _z , M _zi , and M _uz can be obtained from the topic z of each purchase estimated in step 230. U represents the total number of users.

  Step 250) The recommended item calculation unit 20 calculates the recommended item for each user u and each product i from P (i | u) obtained in Step 240 as follows.

ここでNはアイテムの種類数、N_uはユーザuが購買したアイテムの種類数を表す。

Where N is the number of kinds of items, N _u denotes the number of kinds of items for the user u is purchased.

  Step 260) The recommended item calculation unit 20 obtains the recommended item in the recommended item table 73 for each user u and each product i in step 250 in the row where the user ID field value is u and the product ID field value is i. Update to (y = 1 | u, i).

  Next, the processing of step 230 in FIG. 10 will be described in detail.

  FIG. 11 is a flowchart of detailed processing of step 230 in FIG. 10 according to the embodiment of the present invention.

Step 231) The recommended term calculation unit 20 initializes each purchase topic z _m (m = 1,..., M) with an integer random number having a minimum value 1 and a maximum value Z.

Step 232 is repeated a sufficient number of times (eg, 500 times). The number of repetitions may be an arbitrary number set in advance, or may be terminated when the topic z of each purchase converges (when most z _m does not change).

Step 232) The recommended item calculation unit 20 updates the topic z _m (m = 1,..., M) of each purchase.

  Next, the processing of step 232 in FIG. 11 will be described in detail.

  FIG. 12 is a flowchart of detailed processing in step 232 of FIG. 11 according to the embodiment of this invention.

  Steps 2321 and 2322 are performed for all purchases m = 1,.

Step 2321) recommendation term calculation unit 20, each topic candidate k = 1 of該購purchase topic z _m, ..., is calculated by the equation (9) the sampling probability of being assigned to Z.

ここで

here

は該購買を除いたときの回数もしくは変数を表す。つまり、例えば、

Represents the number of times or variables when the purchase is excluded. That is, for example,

は更新前にｚ_mがkにアサインされていれば

If z _m is assigned to k before update

を表し、そうでなければ

Represents otherwise

を表す。また、u_m、i_mはそれぞれm番目の購買履歴のユーザＩＤと商品ＩＤを表す。

Represents. Further, it represents a u _m, i _m is the user ID and product ID of the m-th purchase history, respectively.

Step 2322) recommendation term calculation unit 20, each topic candidate k = 1 topic z _m of該購purchase calculated in Step 2321, ..., based on the sampling probability of being assigned to the K, by generating a random number, z _m Update. Specifically, for k = 2,..., K, the sampling probability is updated according to Equation (10).

さらに、ｋ＝１,…,Kについて、サンプリングを式(11)のように更新する。

Further, for k = 1,..., K, the sampling is updated as in Expression (11).

その後、最小値０、最大値１の小数の乱数を発生させてその値をサンプリング確率が上回る最小のkにアサインする。例を挙げて説明すると、今K=3として、式(9)に基づき算出されたサンプリング確率がそれぞれP(ｚ_m=1)=1.6、P(ｚ_m=2)=0.4、P(ｚ_m=3)=2.0であったとする。つまり、それぞれのトピックに4:1:5の割合でサンプリングされるようにしたい。このとき、式(10)と式(11)を用いてP(ｚ_m=1)=0.4、P(ｚ_m=2)=0.5、P(ｚ_m=3)=1.0とサンプリング確率を更新し、最小値０、最大値１の小数の乱数を発生させてその値をサンプリング確率が上回る最小のｋにアサインすればよい。

Thereafter, a decimal random number having a minimum value of 0 and a maximum value of 1 is generated, and that value is assigned to the minimum k having a higher sampling probability. For example, assuming that K = 3 and the sampling probabilities calculated based on the equation (9) are P (z _m = 1) = 1.6, P (z _m = 2) = 0.4, P (z _m = 3) = 2.0. In other words, we want to sample each topic at a 4: 1: 5 ratio. At this time, using (10) and (11), update the sampling probability to P (z _m = 1) = 0.4, P (z _m = 2) = 0.5, P (z _m = 3) = 1.0. A random number with a minimum value of 0 and a maximum value of 1 may be generated and assigned to the minimum k that has a higher sampling probability.

  Next, processing at the time of calculating a diversification term in an embodiment of the present invention will be described. In this embodiment, this processing is performed when the system administrator uses the external device 4 to issue a diversification term calculation command to the system. You may enable it to set up.

  FIG. 13 is a flowchart of the process when calculating the diversification term in the embodiment of the present invention.

  Step 310) The diversification term calculation unit 30 acquires all the product sets N = {1,..., N} appearing in the purchase history table 71.

  Step 320) The multivariate term calculation unit 30 calculates a popularity term for each product of all the product sets N acquired in Step 310, and updates the popularity term table 731.

  Step 330) The multivariate term calculation unit 30 calculates dissimilarity terms for all the two product pairs in the total product set N acquired in Step 310, and updates the dissimilarity term table 732.

  Step 340) The multivariate term calculation unit 30 calculates similarity terms for all two product pairs in the total product set N acquired in Step 310, and updates the similarity term table 733.

  Next, the process of step 320 in FIG. 13 will be described in detail.

  FIG. 14 is a flowchart of detailed processing of step 320 in FIG. 13 according to the embodiment of the present invention.

Step 321) The diversification term calculation unit 30 refers to the purchase history table 71 and determines the total purchase number M.
To get.

Step 322) diversification term calculation unit 30 refers to the purchase history table 71, obtains the purchase number M _i of each product i in the product set N.

Step 323) For each product i in the product set N, the diversification term calculation unit 30 uses the total purchase number M acquired in Step 321 and the purchase number M _{i of} i acquired in Step 322 as the following formula. The degree term P (y = 1 | i) is calculated.

ステップ３２４）多様化項算出部３０が、商品集合N中の各商品iについて、人気度項テーブル７３１中の商品ＩＤがiの行の人気度フィールドの値をステップ３２３で算出した人気度P(y=1|i)に更新する。

Step 324) For each product i in the product set N, the diversification term calculation unit 30 calculates the popularity P () calculated in step 323 for the value of the popularity field in the row with the product ID i in the popularity term table 731. Update to y = 1 | i).

  Next, the process of step 330 in FIG. 13 will be described in detail.

  FIG. 15 is a flowchart of detailed processing of step 330 in FIG. 13 according to the embodiment of the present invention.

Step 331) The diversification term calculation unit 30 refers to the purchase history table 71 and acquires the number of users U _i who have purchased i for each product i in the product set N.

  Step 332) The diversification term calculation unit 30 refers to the purchase history table 71 and has purchased i's for all the two product pairs (i, i ') in the product set N, although i have been purchased. Number of users never

を取得する。

To get.

Step 333) The diversification term calculation unit 30 calculates the dissimilarity P (y ′ = 0 | i ′, y = 1) between i and i ′ for all the two product pairs (i, i ′) in the product set N. , i) acquired in step 331 and U _i acquired in step 331

から下式のように算出する。

Is calculated as follows.

ここでδはスムージングパラメータであり、予め与えておいた０以上の小数の定数により設定される。

Here, δ is a smoothing parameter, and is set with a decimal constant of 0 or more given in advance.

  Step 334) The diversification term calculation unit 30 sets the product ID in the dissimilarity term table 732 for the pair (i, i ′) of all the two products in the product set N and the upper product ID is i ′. Is updated to the dissimilarity P (y ′ = 0 | i ′, y = 1, i) calculated in step 333.

  Next, the process of step 340 in FIG. 13 will be described in detail.

  FIG. 16 is a flowchart of detailed processing of step 340 in FIG. 13 according to the embodiment of the present invention.

  Step 341) The diversification term calculation unit 30 refers to the purchase history table 71, and the number of users who have not purchased i for each product i in the product set N.

を取得する。

To get.

  Step 342) With reference to the purchase history table 71, the diversification term calculation unit 30 purchases i or purchases i 'for all the two product pairs (i, i') in the product set N. No users

を取得する。

To get.

  Step 343) The diversification term calculation unit 30 determines the similarity P (y ′ = 0 | i ′, y = 0, i) of i and i ′ for all the two product pairs (i, i ′) in the product set N. i) obtained in step 341

とステップ３４２で取得した

And obtained in step 342

から下式のように算出する。

Is calculated as follows.

ここでγはスムージングパラメータであり、予め与えておいた０以上の小数の定数により設定される。

Here, γ is a smoothing parameter, and is set by a decimal constant of 0 or more given in advance.

  Step 344) The diversification term calculation unit 30 sets the product ID in the dissimilarity term table 733 for the pair (i, i ′) of all the two products in the product set N and the upper product ID is i ′. The value of the similarity field is updated to the similarity P (y ′ = 0 | i ′, y = 0, i) calculated in step 343.

  Next, the processing at the time of recommendation history processing in an embodiment of the present invention will be described in detail. In this embodiment, whenever a user who has logged in to the EC site moves from the page on which the recommendation list is posted to another page, the recommendation history is obtained from the linked EC site via the network 3 and the communication unit 90. This process is executed.

  FIG. 17 is a flowchart of processing during recommendation history processing according to an embodiment of the present invention.

  Step 410) The recommendation history processing unit 40 enters the recommendation history table 74 into the recommendation target user, the user clicks on the link to the page of the product information in the recommendation list, or the user according to the recommended product. Insert a row in which the values of the ID field, click position field, product ID field 1,..., L are set. Here, when the destination moved from the recommendation list posting page is not the product information page in the recommendation list, the value of the click position field is set to 0.

  Next, the process at the time of the diversification parameter estimation in one embodiment of this invention is demonstrated. In this embodiment, this processing is performed when the system administrator uses the external device 4 to issue a diversification parameter estimation command to the system, but it is also automatically executed periodically such as every other day. You may enable it to set up.

  FIG. 18 is a flowchart of processing when diversifying parameters are estimated in one embodiment of the present invention.

  Step 510) The diversification parameter estimation unit 50 refers to the recommendation history table 74 and sets the recommendation history set.

を取得する。

To get.

  Step 520) The diversification parameter estimation unit 50 initializes the diversification parameter α with a random decimal number from 0.0 to 10.0. Although a random number is used in the present embodiment, it may be set with a decimal constant of 0 or more (for example, 1.0) given in advance.

  Step 530) A recommended history set of the diversification parameter α initialized by the diversification parameter estimation unit 50 in step 520

に対する対数尤度

Log-likelihood for

を推薦度項テーブル７２に格納されている推薦度項と、多様化項と、式(6)とを用いて算出する。ここで多様化項

Is calculated using the recommendation degree term stored in the recommendation degree term table 72, the diversification term, and Equation (6). Where diversification term

は人気度項テーブル７３１に格納されている人気度項、非類似度項テーブル７３２に格納されている非類似度項、類似度項テーブル７３３に格納されている類似度項から式(15)のように算出する。

Is obtained from the popularity terms stored in the popularity term table 731, the dissimilarity terms stored in the dissimilarity term table 732, and the similarity terms stored in the similarity term table 733. Calculate as follows.

以下、ステップ５４０〜ステップ５６０を、繰り返す。繰り返しの終了条件としては、本実施形態では繰り返しの回数が十分である（図１８ではＴ回としている）か、もしくは、最大尤度の変化量が十分小さくなった場合（図１８では

Thereafter, steps 540 to 560 are repeated. As the repetition end condition, in the present embodiment, the number of repetitions is sufficient (in FIG. 18, T), or when the maximum likelihood change amount is sufficiently small (in FIG. 18).

θは0.0以上の予めシステム管理者により設定されたパラメータ）、としているが、どちらか一方でも構わない。

θ is a parameter set by the system administrator in advance of 0.0 or more), but either one may be used.

  Step 540) The diversification parameter estimation unit 50 sets a recommendation history set of the current diversification parameter α.

に対する対数尤度の勾配

Log-likelihood slope for

を式(16)のように算出する。

Is calculated as shown in Equation (16).

ここでεは勾配を求めるためのパラメータで、0より大だが限りなく0に近い値（例えば1.0×10^-5）がシステム管理者により予め設定されている。

Here, ε is a parameter for obtaining the gradient, and a value larger than 0 but close to 0 (for example, 1.0 × 10 ⁻⁵ ) is preset by the system administrator.

  Step 550) The diversification parameter estimation unit 50 sets the current diversification parameter α and the recommended history set of the current diversification parameter α obtained in step 540.

に対する対数尤度の勾配

Log-likelihood slope for

から多様化パラメータを下式のように更新する。

The diversification parameters are updated as follows:

ここでηは多様化パラメータαを更新するためのパラメータで、0より大で１より小の値（例えば0.1）がシステム管理者により予め設定されている。

Here, η is a parameter for updating the diversification parameter α, and a value larger than 0 and smaller than 1 (for example, 0.1) is preset by the system administrator.

  Step 560) A recommendation history set of the current diversification parameter α updated by the diversification parameter estimation unit 50 in step 550

に対する対数尤度

Log-likelihood for

を推薦度項テーブル７２に格納されている推薦度項と、多様化項と、式(6)とを用いて算出する。

Is calculated using the recommendation degree term stored in the recommendation degree term table 72, the diversification term, and Equation (6).

  Step 570) The diversification parameter estimation unit 50 stores the diversification parameters obtained by repeating steps 540 to 560 in the diversification parameter recording unit 75.

  Next, a process when creating a recommendation list in an embodiment of the present invention will be described. In the present embodiment, this process occurs every time the logged-in user browses a page on the recommendation list, but in addition, the system administrator uses the external device 4 to issue a recommendation list creation command to the system. A recommendation list recording unit may be provided and stored therein.

  The user who is the target of the recommendation list creation is u.

  FIG. 19 is a flowchart of processing when creating a recommendation list according to an embodiment of the present invention.

Step 610) The recommendation list creation unit 60 refers to the purchase history table 71, and the user u obtains an unpurchased product set _Iu .

  In step 620, items to be arranged from the first place to the L place are determined.

  Step 620) The recommendation list creation unit 60 sets the product set as

中の各商品について推薦スコアを算出し、ｌ位の商品

Calculate the recommendation score for each product in

を決定する。

To decide.

ここで推薦項は、P(y_l=1|u,i_l)は推薦リスト作成部６０が推薦度項テーブル７２を参照することにより取得し、多様化項

Here, P (y _l = 1 | u, i _l ) is obtained by referring to the recommendation degree table 72 by the recommendation list creation unit 60, and the recommendation term is a diversification term.

は推薦リスト作成部６０が人気度項テーブル７３１と、非類似度項テーブル７３２と、類似度項テーブル７３３とを参照し、人気度、非類似度、類似度から式(15)のように算出する。

The recommendation list creation unit 60 refers to the popularity term table 731, the dissimilarity term table 732, and the similarity term table 733, and calculates the popularity, dissimilarity, and similarity as shown in Expression (15). To do.

  Note that the processing of each component of the product recommendation device can be constructed as a program and installed in a computer used as the product recommendation device for execution, or distributed via a network.

  In addition, the constructed program can be stored in a portable storage medium such as a hard disk, a flexible disk, or a CD-ROM, and can be installed or distributed in a computer.

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

1 Product recommendation device 2 External terminal 3 Network 4 External device 10 Purchase history processing unit 20 Recommended term calculation unit 30 Diversification term calculation unit 40 Recommendation history processing unit 50 Diversification parameter estimation unit 60 Recommendation list creation unit 70 Storage unit 71 Purchase history Table 72 Recommended item table 73 Diversification item recording unit 74 Recommendation history table 75 Diversification parameter recording unit 731 Popularity item table 732 Dissimilarity item table 733 Similarity table

Claims (5)

A product diversification recommendation device that uses a user's purchase history and recommendation history in an e-commerce service,
A purchase history acquisition unit that acquires a purchase history of each user ID in the E-Commerce service and stores it in a purchase history storage unit;
Using the acquired purchase history, a recommended item that is a probability of purchasing each product of each user is calculated, and a recommended item calculating unit that is stored in a recommended item storage unit;
Diversification term calculation means for calculating a diversification term that is a ratio of users who have purchased a certain product among users who have not purchased a certain product set using the acquired purchase history;
In the E-Commerce service, a recommendation regarding to which user, what product set is a recommendation list, and which detailed information of which product in the recommendation list has been viewed, or which detailed information of any product has not been browsed A recommendation history acquisition means for acquiring a history and storing the history in a recommendation history storage means;
Using the calculated recommendation term, the calculated diversification term, and the acquired recommendation history, a diversification parameter that is a weight of the diversification term with respect to the recommendation term at the time of recommendation score calculation when creating a recommendation list is calculated Diversification parameter calculation means for storing in the diversification parameter storage means;
A recommendation list creating means for creating a recommendation list by using the acquired purchase history, the calculated recommendation item, the calculated diversification item, the acquired recommendation history, and the calculated diversification parameter; ,
A product diversification recommendation device characterized by comprising: The diversification term calculation means includes
Using the purchase history acquired by the purchase history acquisition means, calculate a popularity term that is a probability that each product is purchased, and store it in a popularity term storage means;
Using the acquired purchase history, for each pair of two products, a dissimilarity term is calculated, which is the proportion of users who have purchased one product but who have not purchased the other product. And a dissimilarity term calculation means for storing in the dissimilarity term storage means,
Using the acquired purchase history, for each pair of two products, a similarity term that is the proportion of users who have not purchased one product among the users who have not purchased one product is calculated. , Similarity term calculation means stored in the similarity term storage means,
Means for calculating a diversification term using the calculated popularity term, the calculated dissimilarity term, and the calculated similarity term;
The product diversification recommendation device according to claim 1, including: In an e-commerce service, a product diversification recommendation method that uses a user's purchase history and recommendation history,
Purchase history storage means for holding a purchase history;
A recommendation item storage means for holding a recommendation item;
Recommendation history storage means for holding a recommendation history;
Diversification parameter storage means for holding diversification parameters;
In an apparatus having a purchase history acquisition means, a recommendation term calculation means, a diversification term calculation means, a recommendation history acquisition means, a diversification parameter calculation means, a recommendation list creation means,
The purchase history acquisition unit acquires a purchase history of each user ID in the E-Commerce service and stores it in the purchase history storage unit;
The recommended item calculating means calculates a recommended item that is a probability of purchasing each product of each user using the acquired purchase history, and stores it in the recommended item storage unit;
The diversification term calculation means calculates a diversification term that is a ratio of users who have purchased a certain product among users who have not purchased a certain product set, using the acquired purchase history. Diversification term calculation step to
In the E-Commerce service, the recommendation history acquisition unit uses which product set as a recommendation list to which user, which product in the recommendation list has been viewed, or detailed information on which product. A recommendation history acquisition step of acquiring a recommendation history regarding whether or not also viewed, and storing it in the recommendation history storage means;
The diversification parameter calculation means uses the calculated recommendation item, the calculated diversification item, and the acquired recommendation history to determine the diversification item for the recommendation item at the time of recommendation score calculation when creating a recommendation list. A diversification parameter calculating step of calculating a diversification parameter which is a weight and storing the diversification parameter in the diversification parameter storage unit;
The recommendation list creating means generates a recommendation list using the acquired purchase history, the calculated recommendation item, the calculated diversification item, the acquired recommendation history, and the calculated diversification parameter. A recommendation list creation step to be created;
Product diversification recommendation method characterized by performing. The apparatus further includes a popularity term storage means, a dissimilarity term storage means, and a similarity term storage means,
In the diversification term calculation step,
Using the purchase history acquired in the purchase history acquisition step, calculate a popularity term that is a probability that each product is purchased, and store it in the popularity term storage means;
Using the acquired purchase history, for each pair of two products, a dissimilarity term is calculated, which is the proportion of users who have purchased one product but who have not purchased the other product. And a dissimilarity term calculation step for storing in the dissimilarity term storage means;
Using the acquired purchase history, for each pair of two products, a similarity term that is the proportion of users who have not purchased one product among the users who have not purchased one product is calculated. , A similarity term calculation step stored in the similarity term storage means;
Calculating a diversification term using the calculated popularity term, the calculated dissimilarity term, and the calculated similarity term;
The product diversification recommendation method according to claim 3. Computer
A product diversification recommendation program for functioning as each unit of the product diversification recommendation device according to claim 1.

Images