JP5654977B2 - Product recommendation device, method and program - Google Patents

Description

  The present invention relates to a product recommendation device, method, and program, and more particularly, to a product recommendation device, method, and program for user product purchase provided by an e-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 users with similar preferences Has been proposed (see, for example, Non-Patent Document 1).

  Similarly to the above method, the account is evaluated for each piece of information, and the text included in the information evaluated by the account in the past evaluation history is focused. A method has been proposed in which the text reflects the interest of the account and recommends information in which words similar to the text appear (see, for example, Non-Patent Document 2).

  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).

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. Mooney R.J. et al, "Content-Based Book Recommending Using Learning for Text Categorization", Proceedings of the 5th ACM Conference on Digital Libraries, ACM Press New York, 2000

  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, it is impossible to recommend a product that has not been purchased by anyone. For this reason, only popular and famous products tend to be recommended, and the user often already knows the product, and recommending it does not lead to a new purchase.

  In addition, when a product is recommended by the method described in Non-Patent Document 2, a word not related to the purchase preference of the user, such as “people” or “information” in Japanese, becomes noise and the user buys it. There is a problem that products with high probability cannot be recommended well.

  The present invention has been made in view of the above points. In the E-Commerce service, the purchase history is estimated by estimating which word in the metadata attached to the product from the purchase history is related to the purchase. It is an object of the present invention to provide a product recommendation device, method and program that can recommend a product that a user has a high probability of buying.

In order to solve the above problems, the present invention (Claim 1) is a product recommendation device that uses product information and user purchase information in an E-Commerce service,
Product information acquisition means for acquiring product information in the E-Commerce service and storing the product information in the product information storage means;
Purchase information acquisition means for acquiring purchase information of each user ID in the E-Commerce service, and storing the purchase information in the purchase information storage means;
Using the information on whether or not each user has purchased each product in the purchase information in the purchase information storage means, and the metadata assigned to each product in the product information in the product information storage means, each user User topic estimation means for estimating purchase topics of
The relevance of each word to the purchase information is estimated by using information on whether or not each user has purchased each product in the purchase information and metadata given to each product in the product information. Relevance estimation means;
The acquired purchase was whether information about each product for each user in the purchase information, with reference to the metadata assigned to each product in the product information, and word topic estimating means for estimating a word topic ,
Using the purchase topic of each user , the relevance of each word to the purchase information, and the word topic, the probability that each user will purchase each product is calculated, and the recommendation score stored in the recommendation score storage means A calculation means;
Given a user identifier, for each product stored in the product information storage means, a recommendation score stored in the recommendation score storage means is obtained, and the top N products of the recommendation score are obtained. Product recommendation means for selecting an identifier as a recommended product.

  In the present invention (Claim 2), the user topic estimation unit, the relevance estimation unit, and the word topic estimation unit use Gibbs sampling.

  According to the present invention, the present invention uses both the purchase history and the product metadata, and which word in the product metadata is related to the user's purchase, which was impossible with the prior art. Can be estimated. Further, by using the relevance to the purchase of words, it is possible to recommend to a user with high accuracy even if the product has never been purchased. This leads to an increase in purchasing in e-commerce services.

It is a block diagram of the merchandise information recommendation apparatus in one embodiment of this invention. It is an example of the goods information table in one embodiment of this invention. It is an example of the purchase information table in one embodiment of this invention. It is an example of the recommendation score table in one embodiment of this invention. It is a flowchart of the process at the time of the merchandise information update in one embodiment of this invention. It is a flowchart of the process at the time of purchase in one embodiment of this invention. It is a flowchart of the process at the time of the recommendation score update in one embodiment of this invention. It is a flowchart of the detailed process of step 350 of FIG. 7 in one embodiment of this invention. It is a flowchart of the detailed process of step 354 of FIG. 8 in one embodiment of this invention. It is a flowchart of the detailed process of step 355 of FIG. 8 in one embodiment of this invention. It is a flowchart of the detailed process of step 356 of FIG. 8 in one embodiment of this invention. It is a flowchart of the process at the time of the product recommendation in one embodiment of this invention.

  Embodiments of the present invention will be described below with reference to the drawings.

  An object of the present invention is to recommend a product that has a high probability of being purchased by a user even if it has never been purchased in an E-Commerce service. Specifically, using a purchase history, it is estimated which word is related to purchase. After that, the user's interest is estimated using the purchase history and the relevance of the word to the purchase, and a method of recommending a product that matches the user's interest is realized. Therefore, even a product that has never been purchased can be recommended as long as the product has metadata including words related to the user's purchase.

  In the present invention, the purchase history and the generation of the product metadata are associated and modeled. Now, the purchase history is represented by {i, u}.

はどの商品が購買されたかを表し、

Indicates which product was purchased,

はどのユーザにより購買されたかを表す。ただし、Ｍは全購買数を表す。同様に、商品のメタデータを｛i，w｝で表す。

Represents by which user the purchase was made. However, M represents the total number of purchases. Similarly, merchandise metadata is represented by {i, w}.

はどの商品へのメタデータなのかを、

Is the product's metadata,

はどの語がメタデータとして付与されたかを表す。但し、Nは全商品のメタデータに含まれる語数の合計である。このとき、｛ｉ，ｕ，ｊ，ｗ｝が生成される確率が以下の式(1)のように展開できると仮定する。

Represents which word is given as metadata. N is the total number of words included in the metadata of all products. At this time, it is assumed that the probability that {i, u, j, w} is generated can be expanded as shown in the following equation (1).

ここで、

here,

は、購買のトピック（非特許文献１でいうところのユーザのpreference）を、

Is the purchasing topic (user preference in Non-Patent Document 1)

は、当該語の購買への関連性をそれぞれ表す。

Represents the relevance of the term to purchase, respectively.

  Let's use a concrete example. There is a purchase history consisting of several users and several products, and each product contains metadata text. Suppose that there is a product with metadata such as "Great", "SF", "Horror", "Work". From the purchase history, only “SF” is determined to be a word related to the recommendation. As a result, even if the product has never been bought, this product is recommended to users who like “SF”.

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

  FIG. 1 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 product information processing unit 10, a purchase information processing unit 20, a recommendation score calculation unit 30, a product recommendation unit 40, a recording unit 50, a communication unit 60, and an input / output unit 70. The unit 60 is connected to the network 3, and the input / output unit 70 is connected to the external device 4 such as an input device or a display device.

  The recommendation score calculation unit 30 includes a user topic estimation unit 31, a relevance estimation unit 32, and a word topic estimation unit 33.

  The storage unit 50 includes a product information table 51, a purchase information table 52, and a recommendation score table 53. Each table will be described below.

<Product information table 51>
As shown in FIG. 2, the product information table 51 includes a product ID field and a metadata field. The product ID field is an identifier that identifies the product, and is set by the product information processing unit 10. In the metadata field, the metadata of the product is set as text data by the product information processing unit 10. This metadata may be anything as long as it is in the form of text, such as the text of a book, a product introduction summary created by a content distributor, and a tag attached by a user.

<Purchase information table 52>
As shown in FIG. 3, the purchase information table 52 includes a product ID field and a user ID field. The product ID field corresponds to the identifier of the product that appears in the product information table, and is set by the purchase information processing unit 20. In the user ID field, an identifier for identifying a user who has purchased the product is set by the purchase information processing unit 20.

<Recommended score table 53>
As shown in FIG. 4, the recommendation score table 53 includes a user ID field, a product ID field, and a recommendation score field. In the recommendation score field, a recommendation score P (i, u) for the product i of the account u is set by the recommendation score calculation unit 30.

  First, the process at the time of the merchandise information update in one embodiment of this invention is demonstrated in detail. The product information is updated by the system administrator in the case where the metadata is provided from a service provider such as the text of a book or a content introduction. If the metadata is provided by a user such as a social bookmark tag, both the system administrator and the user may cause this processing.

  FIG. 5 is a flowchart of the process when updating the product information according to the embodiment of the present invention. The product information update timing may be managed manually by a system administrator, for example. Assume that information about the product i has been updated.

  Step 110) The product information processing unit 10 configures a line in which the information on the product i is written in the product information table 51. If the information on the product i is already included in the product information table 51, the information is updated as such. If not, insert a new line.

  FIG. 6 is a flowchart of processing at the time of purchase in one embodiment of the present invention.

  Step 210) The purchase information processing unit 20 inserts in the purchase information table 52 a line in which the value of the product ID field and the user ID field is set according to the purchased product and the purchased user.

  FIG. 7 is a flowchart of a process at the time of updating a recommendation score according to an embodiment of the present invention. The timing for updating the recommendation score may be manually managed by a system administrator, for example, or may be automatically performed periodically.

  Step 310) The recommendation score calculation unit 30 calculates the number of topics 推薦 which is a parameter for recommendation, a prior parameter ξ, P (i | z), and P (j | y) for P (z) calculation. Prior parameter β for calculating prior parameter α, P (u | z), prior parameter γ for calculating P (w | y, r), prior parameter δ for calculating P (c), P (r | c) Set a prior parameter η for calculation. These parameters are set by constants given in advance.

  Step 320) The recommended score calculation unit 30 refers to the purchase information table 52 and acquires the user u and the product i for each purchase.

  Step 330) The recommended score calculation unit 30 refers to the product information table 51 and acquires product metadata.

  Step 340) The recommendation score calculation unit 30 divides the text in the product metadata acquired in Step 330 into words by a technique such as morphological analysis, and item j, word to which each word is assigned Get w.

  Step 350) The recommendation score calculation unit 30 uses the Collapsed Gibbs sampling described in Reference 1 (Bishop, MC, “Pattern recognition and machine learning”, 2006) to obtain the topic z of each purchase and the topic y of each word in the metadata. The relevance r with the purchase of each word in the metadata is estimated.

  Step 360) The parameter Κ, ξ, α, β, γ, δ, η set by the recommended score calculation unit 30 in Step 310, the topic z of each purchase estimated in Step 350, and each word in the metadata The recommendation score P (i, u) for each product i of each user u is calculated using the relationship r between the topic y and the purchase of each word in the metadata and the following equation (2).

ここでM_kはトピックkにアサインされた購買数、Ｎ_kはトピックkにアサインされたメタデータ中の語の総数、M_kiはトピックkにアサインされ、かつ商品iに関する購買の数、N_kiは商品iのメタデータ中の語でトピックｋにアサインされたものの数、M_kuはトピックkにアサインされたユーザｕの購買数を表す。M_kとM_kuとM_ki、M_kuはステップ３５０にて推定した各購買のトピックｚにより求めることができ、Ｎ_kならびにN_kiはメタデータ中の各語の購買との関連性ｒを用いて推定されたメタデータ中の各語のトピックｙにより求めることができる。なお、Ｕはユーザの総数を表す。

Where M _k is the number of purchases assigned to topic k, N _k is the total number of words in the metadata assigned to topic k, M _ki is the number of purchases assigned to topic k and for item i, N _ki Is the number of words in the metadata of the product i assigned to the topic k, and M _ku represents the number of purchases of the user u assigned to the topic k. M _k , M _ku , M _ki , and M _ku can be obtained from the topic z of each purchase estimated in step 350, and N _k and N _ki use the relationship r with the purchase of each word in the metadata. The topic y of each word in the estimated metadata can be obtained. U represents the total number of users.

  Step 370) The recommendation score calculation unit 30 sets a recommendation score P (i, u) for each user u and each product i in the recommendation score table 53 where the value of the user ID field is u and the value of the product ID field is i. Update to

  Next, the process of step 350 in FIG. 7 will be described in detail.

  FIG. 8 is a flowchart of detailed processing of step 350 in FIG. 7 according to the embodiment of the present invention.

Step 351) The recommendation score calculation unit 30 initializes each purchase topic z _m (m = 1,..., M) with a minimum value 1 and an integer K of the maximum value.

Step 352) The recommendation score calculation unit 30 initializes the topic y _n (n = 1,..., N) of each word in the metadata with the minimum value 1 and the integer K of the maximum value.

Step 353) The recommendation score calculation unit 30 initializes the relevance r _n (n = 1,..., N) of each word in the metadata with the minimum value 0 and the maximum value integer 1.

  Steps 354 to 356 are repeated a sufficient number of times (for example, 500 times). The number of repetitions may be an arbitrary number set in advance. If the relevance r between the topic z of each purchase, the topic y of each word in the metadata, and the purchase of each word in the metadata converges, it is terminated there. You may do it.

Step 354) The user topic estimation unit 31 sets each purchase topic z _m (m = 1,..., M).
Update.

Step 355) The relevance estimation unit 32 relevance of each word in the metadata to the purchase r _n (n = 1,..., N)
Update.

Step 356) The word topic estimation unit 33 determines the topic y _n (n = 1,..., N) of each word in the metadata.
Update.

  Next, the process of step 354 in FIG. 8 will be described in detail.

  FIG. 9 is a flowchart of detailed processing of step 354 in FIG. 8 according to the embodiment of the present invention.

  Steps 3541 and 3542 are performed for all purchases m = 1,.

Step 3541) The user topic estimation unit 31 calculates the sampling probability assigned to each topic candidate k = 1,..., K of the purchased topic z _m by the following equation (3).

ここで＼ｍ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

Here, \ m represents the number of times or variable 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。

Represents.

Step 3542) The user topic estimating unit 31, each topic candidate k = 1 topic z _m of該購purchase calculated in Step 3541, ..., 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 (4).

さらに、k＝１，…，Kについて、サンプリングを式（5）のように更新する。

Further, for k = 1,..., K, the sampling is updated as in Expression (5).

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（3）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(4)と式(5)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, the sampling probabilities calculated based on Equation (3) 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 (4) and (5), 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, the processing of step 355 in FIG. 8 will be described in detail.

  FIG. 10 is a flowchart of detailed processing of step 355 in FIG. 8 according to the embodiment of the present invention.

  Steps 3551 and 3552 are performed for all words n = 1,..., N in the metadata.

Step 3551) The relevance estimation unit 32 calculates the sampling probability that the relevance r _n to the purchase of the word is assigned to 0 or 1 by the equations (6) and (7).

ここでN'はメタデータ中の語数、N'₀は購買への関連性が０であるメタデータ中の語数を、N'_0wは購買への関連性が０であるメタデータ中の語wの出現数を、N'_yは購買への関連性が１で、トピックyにアサインされたメタデータ中の語数を、N'_ywは購買への関連性が１で、トピックyにアサインされたメタデータ中の語wの出現数を、Wはメタデータの語彙数をそれぞれ表す。＼nは該語を除いたときの回数もしくは変数を表す。

Here, N ′ is the number of words in the metadata, N ′ ₀ is the number of words in the metadata where the relevance to purchase is 0, and N ′ _0w is the word w in the metadata where the relevance to purchase is 0. , N ' _y is the relevance to purchase with 1 and the number of words in the metadata assigned to topic y, N' _yw is the relevance to purchase with 1 and assigned to topic y W represents the number of occurrences of the word w in the metadata, and W represents the number of words in the metadata. \ N represents the number of times or variables when the word is removed.

Step 3552) The relevance estimation unit 32 generates a random number based on the sampling probability of the relevance r _n to the purchase of the word calculated in Step 3551, and updates r _n . Details of this processing are the same as in step 3542.

  Next, the processing in step 356 in FIG. 8 will be described in detail.

  FIG. 11 is a flowchart of detailed processing of step 356 in FIG. 8 according to the embodiment of this invention.

  Steps 3561 and 3562 are performed for all words n = 1,..., N in the metadata.

Step 3561) The word topic estimation unit 33 calculates a sampling probability that the topic y _{n of the} word is assigned to each of the topic candidates k = 1,. If the relevance of the word updated in step 355 to purchase is 0, equation (8) is used, and if it is 1, equation (9) is used.

ステップ３５６２）ワードトピック推定部３３が、ステップ３５６１にて算出した該語のトピックｙ_nのサンプリング確率に基づき、乱数を発生させて、ｙ_nを更新する。本処理の詳細はステップ３５４２と同様である。

Step 3562) word topic estimating unit 33 based on the sampling probability topic y _n of word or calculated in step 3561, by generating a random number, updating y _n. Details of this processing are the same as in step 3542.

  FIG. 12 is a flowchart of processing at the time of product recommendation according to an embodiment of the present invention. For example, this process may be executed when the user logs in to the system.

  Step 410) The merchandise information processing unit 10 refers to the merchandise information table 51 and acquires a merchandise set.

  Step 420) The product recommendation unit 40 refers to the recommendation score table 53 for each product i included in the product set, obtains a recommendation score P (i, u) for the product i in the account u, and stores it in the account u. Decide which products to recommend. At that time, an arbitrary number K (K = 1, 2, 3,...) Preset in order from the largest recommended score P (i, u) may be set as a product to be recommended. A certain threshold may be set for (i, u), and a product whose recommendation score exceeds that value may be recommended.

  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 Product information processing unit 20 Purchase information processing unit 30 Recommended score calculation unit 31 User topic estimation unit 32 Relevance estimation unit 33 Word topic estimation unit 40 Product recommendation unit 50 Storage unit 51 Product information table 52 Purchasing information table 53 Recommended score table 60 Communication unit 70 Input / output unit

Claims (5)

In an electronic commerce (E-Commerce) service, a product recommendation device that uses product information and user purchase information,
Product information acquisition means for acquiring product information in the E-Commerce service and storing the product information in the product information storage means;
Purchase information acquisition means for acquiring purchase information of each user ID in the E-Commerce service, and storing the purchase information in the purchase information storage means;
Using the information on whether or not each user has purchased each product in the purchase information in the purchase information storage means, and the metadata assigned to each product in the product information in the product information storage means, each user User topic estimation means for estimating purchase topics of
The relevance of each word to the purchase information is estimated by using information on whether or not each user has purchased each product in the purchase information and metadata given to each product in the product information. Relevance estimation means;
The acquired purchase was whether information about each product for each user in the purchase information, with reference to the metadata assigned to each product in the product information, and word topic estimating means for estimating a word topic ,
Using the purchase topic of each user , the relevance of each word to the purchase information, and the word topic, the probability that each user will purchase each product is calculated, and the recommendation score stored in the recommendation score storage means A calculation means;
Given a user identifier, for each product stored in the product information storage means, a recommendation score stored in the recommendation score storage means is obtained, and the top N products of the recommendation score are obtained. Product recommendation means for selecting an identifier as a recommended product,
A product recommendation device comprising: The product recommendation device according to claim 1, wherein the user topic estimation unit, the relevance estimation unit, and the word topic estimation unit use Gibbs sampling. A product recommendation method using product information and user purchase information in an electronic commerce service (E-Commerce) service,
A product information acquisition unit acquires product information in the E-Commerce service, and stores the product information in the product information storage unit;
Purchasing information acquisition means acquires purchasing information for each user ID in the E-Commerce service, and stores the purchasing information in the purchasing information storage means;
Information about whether or not the user topic estimation unit has purchased each product of each user in the purchase information in the purchase information storage unit, and metadata given to each product in the product information in the product information storage unit A user topic estimation step for estimating the purchase topic of each user using
The relevance estimating means uses the information on whether or not each user has purchased each product in the purchase information and the metadata assigned to each product in the product information, to purchase information of each word. A relevance estimation step for estimating the relevance of
Word topic estimating means, by using the purchase was whether information about each product for each user in the purchase information, and meta data attached to each product in the product information, words to estimate the word topic A topic estimation step;
The recommendation score calculation means calculates the probability that each user purchases each product using the purchase topic of each user , the relevance of each word to the purchase information, and the word topic, and stores a recommendation score. A recommendation score calculation step stored in the means;
Commodity recommendation means, given the identifier of the user, for each commodity stored in the product information storage unit, acquires the recommendation score stored in said recommendation score storage means, the upper of the recommendation score A product recommendation step of selecting identifiers of N products as recommended products;
A method for recommending products. The product recommendation method according to claim 3, wherein Gibbs sampling is used in the user topic estimation step, the relevance estimation step, and the word topic estimation step. Computer
The product recommendation program for functioning as each means of the product recommendation apparatus of Claim 1 or 2.

Images