JP2017215647A - Selection device, selection method and selection program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent increase in calculation cost when estimating evaluation of a user.SOLUTION: A selection device 10 comprises: an acquisition part which acquires a fourth matrix M30 in which the component value is set in such a manner that a product of a first matrix M20, the fourth matrix M30 and a second matrix M40 becomes a third matrix M10, with respect to the fourth matrix M30 calculated from the first matrix M20 including information indicating features of a user, the second matrix M40 including information indicating features of a proposal object and the third matrix M10 including information indicating evaluation of the user with respect to the proposal object; a retrieval part which defines information indicating features of a user to be a proposal destination of the proposal object as a query, defines a value of a product of the fourth matrix M30 and the second matrix M40 as a score of each proposal object and retrieves a predetermined number of proposal objects successively from a highest value among inner product values that are calculated from the query and the scores; and a selection part which selects a proposal object to be proposed to the user on the basis of a result of the retrieval by the retrieval part.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a selection device, a selection method, and a selection program.

  In recent years, with the rapid spread of the Internet, information distribution via the Internet has been actively performed. As an example of such information distribution, proposals are made to users from various contents such as voice, moving images, news, blogs, etc. that are to be distributed and products that are subject to transactions in online shopping malls, etc. 2. Description of the Related Art A technique for selecting a target to be performed (hereinafter referred to as “proposed target”) and notifying a user of the selected proposed target is known.

  As a method for selecting such a proposal target, a method called MF (Matrix Factorization) is known. For example, a server that selects a proposal target using MF generates two latent matrices that reproduce an evaluation value matrix from an evaluation value matrix that includes a user's evaluation of some contents, products, and the like. In addition, the server estimates the evaluation of the user for the unevaluated proposal target by recalculating the evaluation value matrix from the two generated latent matrices. Then, the server selects a proposal target based on the estimation result.

JP 2013-218485 A Japanese Patent Laying-Open No. 2015-060330

  However, in the above-described conventional technology, there is a risk that the calculation cost for estimating the user's evaluation increases.

  For example, in the above-described conventional technology, when a user makes a new evaluation, when the user changes the evaluation, when a new user is added, or content or a product to be proposed is added In some cases, the latent matrix is recalculated every time the evaluation value matrix that is the source of the latent matrix changes. As a result, in the prior art, the calculation cost associated with the calculation of the latent matrix increases.

  The present application has been made in view of the above, and an object thereof is to prevent an increase in calculation cost when estimating a user's evaluation.

  The selection device according to the present application includes a first matrix including information indicating characteristics of a user, a second matrix including information indicating characteristics of the proposal target, and information indicating evaluation of the user with respect to the proposal target. A fourth matrix calculated from three matrices, wherein the values of the components are set such that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix And an information indicating the feature of the user who is the proposal target of the proposal target as a query, and the value of the product of the fourth matrix and the second matrix as the proposal target for the feature of the user A search unit that searches for a predetermined number of proposal targets from a higher value of the inner product value calculated from the query and the score, and for the user based on a search result by the search unit A selection unit for selecting a proposal target to be proposed And features.

  According to one aspect of the embodiment, it is possible to prevent an increase in calculation cost when estimating a user's evaluation.

FIG. 1 is a diagram illustrating an example of a selection process executed by the selection device according to the embodiment. FIG. 2 is a diagram illustrating a configuration example of the selection device according to the embodiment. FIG. 3 is a diagram illustrating an example of information registered in the product database according to the embodiment. FIG. 4 is a diagram illustrating an example of information registered in the evaluation value database according to the embodiment. FIG. 5 is a diagram illustrating an example of information registered in the user database according to the embodiment. FIG. 6 is a diagram illustrating an example of information registered in the latent element database according to the embodiment. FIG. 7 is a diagram illustrating an example of a process in which the selection device according to the embodiment calculates a latent element matrix. FIG. 8 is a diagram illustrating an example of processing executed when a new user is added. FIG. 9 is a diagram illustrating an example of processing executed when a product is added. FIG. 10 is a first diagram illustrating an example of processing for selecting a product with a high evaluation value using the WAND algorithm. FIG. 11 is a second diagram illustrating an example of processing for selecting a product having a high evaluation value using the WAND algorithm. FIG. 12 is a third diagram illustrating an example of processing for selecting a product having a high evaluation value using the WAND algorithm. FIG. 13 is a flowchart illustrating an example of a flow of processing in which the selection device according to the embodiment generates a latent element matrix. FIG. 14 is a flowchart illustrating an example of the flow of selection processing executed by the selection device according to the embodiment. FIG. 15 is a diagram illustrating an example of a hardware configuration.

  Hereinafter, a mode for implementing a selection device, a selection method, and a selection program according to the present application (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings. In addition, the selection apparatus, the selection method, and the selection program which concern on this application are not limited by this embodiment. In the following embodiments, the same portions are denoted by the same reference numerals, and redundant description is omitted.

[Embodiment]
[1. Example of selection device)
First, an example of selection processing executed by the selection device will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of a selection process executed by the selection device according to the embodiment. In FIG. 1, the selection device 10 can communicate with a user terminal 100 used by a user U01 via a predetermined network N such as the Internet.

  The user terminal 100 is a smart device such as a smartphone or a tablet, and is a mobile terminal device that can communicate with an arbitrary server device via a wireless communication network such as 3G (3rd Generation) or LTE (Long Term Evolution). It is. Note that the user terminal 100 may be not only a smart device but also an information processing apparatus such as a desktop PC or a notebook PC.

  The selection device 10 is an information processing device that distributes various web pages to the user terminal 100, and is realized by, for example, a server device or a cloud system. A more specific example will be described. The selection device 10 is an information processing device that provides various services related to an online shopping mall. For example, the selection device 10 is generically referred to as a product or service (hereinafter, “product”) that is an object of electronic commerce. )) And various web pages that accept purchases of products.

  The selection device 10 may be a distribution device that distributes various contents such as music and moving images. That is, the selection device 10 may perform distribution of arbitrary information as long as information selection processing for distributing information to the user terminal 100 is executed.

  Here, the selection device 10 makes various recommendations to the user. For example, the selection device 10 selects a product that the user is expected to like based on the user's evaluation of each product, and information that suggests the purchase of the selected product, that is, the user is expected to like the product. Deliver advertisements for products to users.

[1-1. About MF]
Here, as a method for selecting information to be proposed to the user, a method called MF is known. Hereinafter, an example of processing executed by the server device that selects information to be proposed to the user using MF will be described. In the following description, an example of processing for selecting content to be proposed to the user from content such as moving images will be described.

  For example, the server device acquires an evaluation value matrix R indicating the user's evaluation for each content. More specifically, the server device is a matrix in which each user is associated with each row and each content is associated with each column, and an evaluation value matrix R in which evaluation values of each user for each content are stored. To get. In general, since not all users are evaluating all contents, such an evaluation value matrix R is a sparse matrix in which only some components are input.

  Here, the server device decomposes the evaluation value matrix R into two latent matrices. For example, when the evaluation value matrix R is a Nu × Nj matrix, the server device sets a latent matrix P that is a Nu × k matrix and a latent matrix Q that is a k × Nj matrix. Further, the server device randomly sets values according to the normal distribution as initial values of the components of the latent matrix P and the latent matrix Q. Then, the server device updates the components of the latent matrix P and the latent matrix Q so that the value of the cost function based on the difference between the product of the latent matrix P and the latent matrix Q and the evaluation value matrix R is minimized. To do. When the value of the cost function becomes sufficiently small, the server device calculates the product of the latent matrix P and the latent matrix Q as a restoration matrix R ′.

  The restoration matrix R ′ calculated in this way includes not only the component input to the evaluation value matrix R but also the component that has not been input to the evaluation value matrix R. That is, the restoration matrix R ′ includes predicted values of user evaluation for each content. Therefore, the server device predicts the user's evaluation for each content using the restoration matrix R ′, and selects the content to be proposed to the user based on the prediction result.

  However, in the MF technique described above, when the user's evaluation for each content changes, that is, when the component of the evaluation value matrix R changes, the latent matrix P that reproduces the changed evaluation value matrix R and the latent matrix The matrix Q needs to be calculated again. Further, when a new user or content is added, the number of rows and the number of columns of the evaluation value matrix R changes, so that a new latent matrix P and latent matrix Q that reproduce the new evaluation value matrix R are obtained. It is necessary to calculate again. As a result, the above-described MF technique may increase the calculation cost. Especially in services where the user's evaluation of content changes instantaneously, such as news recommendations and product recommendations, compared to services such as movie distribution services where the user evaluation of content does not change much. The update frequency of the latent matrix P and the latent matrix Q is increased.

[1-2. About selection processing)
Therefore, the selection device 10 executes the following processing. First, the selection device 10 includes a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the proposal target, and a third matrix including information indicating the evaluation of the user with respect to the proposal target. The fourth matrix calculated from the above is acquired, and the fourth matrix is set such that the product of the first matrix, the fourth matrix, and the second matrix is the third matrix. Further, the selection device 10 uses the information indicating the feature of the user who is the proposal target of the proposal target as a query, and sets the value of the product of the fourth matrix and the second matrix as the score of each proposal target for the user feature, A predetermined number of proposal targets are searched from a higher inner product value calculated from the query and the score. And the selection apparatus 10 selects the proposal object proposed with respect to a user based on the search result by a search part.

[1-3. About IMC (Inductive Matrix Completion)
More specifically, the selection device 10 generates a latent element matrix W that reproduces the evaluation value matrix R using a technique called IMC. For example, the selection device 10 acquires a user feature matrix X in which user features are arranged in the row direction as the first matrix, and arranges information indicating the feature of the product to be proposed in the column direction as the second matrix. The obtained product feature matrix Y is acquired. Moreover, the selection apparatus 10 acquires the evaluation value matrix R which has arrange | positioned the information which shows the user's evaluation with respect to goods in a row direction as a 3rd matrix.

  The user characteristics are various types of information indicating user attributes such as the user's gender, age, type of user terminal 100 used by the user, income, address, hobbies and preferences. . The feature of the product includes arbitrary information indicating an attribute of the product itself such as which category the product belongs to and a price range of the product. Note that the product features may include information indicating the sales method of the product, such as whether it is a sale target product, whether a tax-included price is displayed, and whether the shipping fee is free. .

  The selection device 10 then reproduces the evaluation value matrix R (that is, R≈X × W × Y) so that the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y is reproduced. ), The latent element matrix W is calculated as the fourth matrix. That is, the selection device 10 performs matrix decomposition on the evaluation value matrix R based on the user characteristics and the product characteristics.

  More specifically, the selection device 10 sets a latent element matrix U having the same number of rows as the number of columns of the user feature matrix X and a latent element matrix V having the same number of columns as the number of rows of the product feature matrix Y. The product of the user feature matrix X, the latent element matrix U, the latent element matrix V, and the product feature matrix Y reproduces the evaluation value matrix R (that is, R≈X × U × V × Y). The latent element matrix U and the latent element matrix V are generated using ALS (Alternating Least Squares) or the like. Then, the selection device 10 calculates the product of the generated latent element matrix U and the latent element matrix V as a latent element matrix W. Then, the selection device 10 calculates a restoration matrix R ′ from the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y, and based on the calculated component values of the restoration matrix R ′, Select a product to be proposed to the user.

  The latent element matrix W calculated in this way is not information indicating the user's evaluation for the product, but a latent element for estimating the user's evaluation for the product from the relationship between the user's characteristics and the product characteristics. It becomes the information which shows. For this reason, since the selection device 10 can specify the attributes of a new user or a new product even when a new user is added or a new product is added. If there is, the evaluation value can be calculated without updating the latent element matrix W. As a result, when the latent element matrix W is generated, the selection device 10 prevents an increase in calculation cost for updating the latent element matrix W even when a user or a product is added. I can do it.

[1-4. Regarding calculation of restoration matrix R ′]
Here, when calculating the restoration matrix R ′, the selection device 10 calculates the product of the row in which the feature of the user as the product proposal destination is arranged, the latent element matrix W, and the product feature matrix Y. To do. More specifically, when the selection device 10 receives various requests from a user, such as when a search request for a product is received from a user or when a proposal request for a product is received from a user, The product of the row in which the feature of the user is arranged, the latent element matrix W, and the product feature matrix Y is calculated.

  For example, when the user feature matrix X includes the feature of the user as the proposal destination, the selection device 10 extracts a row including the feature of the user as the proposal destination from the user feature matrix X. Then, the selection device 10 integrates the extracted row, the latent element matrix W, and the product feature matrix Y, so that the restoration matrix R ′ includes a row that includes an evaluation of each product by the user as a proposal destination. calculate.

  On the other hand, when the new user is the proposal destination, that is, when the user feature matrix X does not include the feature of the user who is the proposal destination, the selection device 10 displays the characteristics of the new user. Is generated. Such a row does not have to include all the features that can be acquired. For example, a predetermined initial value may be included as a component. And the selection apparatus 10 calculates the row | line | column containing the evaluation to each goods by the user who becomes a proposal destination by integrating the produced | generated row | line | column, the latent element matrix W, and the product feature matrix Y. The selection device 10 adds the newly generated row to the user feature matrix X, extracts a row indicating the feature of the user who is the proposal destination from the user feature matrix X, and extracts the extracted row and the latent element matrix. By integrating W and the product feature matrix Y, a row including evaluations of each product by the user as a proposal destination may be calculated in the restoration matrix R ′.

  Moreover, the selection apparatus 10 should just add the column which shows the characteristic of a new product to the product feature matrix Y, when a new product is added to the proposal object. Even when such an addition is performed, the selection device 10 integrates the row indicating the feature of the user as the proposal destination, the latent element matrix W, and the product feature matrix Y, thereby restoring the restoration matrix R ′. Among them, it is possible to calculate a row including evaluation of each product by a user who is a proposal destination.

  In this way, the selection device 10 selects a row obtained as a product of the feature of the user as the proposal destination, the latent element matrix W, and the product feature matrix Y as each product by the user as the proposal destination. It is calculated as a row corresponding to the user who is the proposal destination in the restoration matrix R ′. For this reason, the selection device 10 can use the latent element matrix W as it is even when a user is added or a product is added. For this reason, the selection apparatus 10 can prevent an increase in calculation cost when estimating the user's evaluation.

[1-5. About WAND]
Here, when a product that is predicted to be of interest to a user is proposed, the selection device 10 executes the following process. First, the selection device 10 accumulates the row indicating the feature of the user who is the proposal destination, the latent element matrix W, and the product feature matrix Y, thereby indicating a value indicating the evaluation of each product by the user who is the proposal destination. Is calculated. And the selection apparatus 10 selects a predetermined number of goods as a proposal object from the one where the calculated value is higher, and delivers the information regarding the selected goods to a user. However, depending on the number of products targeted for electronic commerce, it may not be possible to quickly search for products targeted for proposal. For example, when a product to be subject to electronic commerce exists in an order of 100 million, the calculation cost for searching for a proposal target increases.

  Therefore, the selection device 10 searches for a product to be proposed using a so-called WAND algorithm. Specifically, the selection device 10 calculates in advance an integration matrix WY that is a product of the latent element matrix W and the product feature matrix Y. Then, the selection device 10 uses the characteristics of the user who is the provision destination as a query, sets the values of each component of the integration matrix WY as the scores of each product for the query, and sequentially determines the values of the inner products of the query and the score in descending order. Select the number of products. That is, the selection device 10 considers the feature of the user who is the proposal destination as a query, and uses the WAND algorithm in which the integration matrix WY is regarded as an index, to select a product that is expected to be more interested by the user. Search for.

  When such a WAND algorithm is adopted, among the inner product values calculated so far, a predetermined number of inner product values are retained as candidates, and based on the query value and the score value, The calculation of the inner product value that does not reach the candidate can be skipped. For this reason, the selection apparatus 10 can reduce the calculation cost at the time of calculating the user's evaluation value with respect to goods. Note that the selection device 10 uses a score for the query to be selected, and if it is a so-called Top-k retry algorithm that selects a predetermined number of products in descending order of the inner product of the query and the score, the WAND algorithm In addition, any algorithm may be adopted.

[1-6. (Applicable selection process)
In the above-described example, the selection device 10 obtains a product feature matrix Y that indicates the characteristics of the product that is the object of electronic commerce, and an evaluation value matrix R that indicates the evaluation of each user for the product that is the object of electronic commerce. Thus, the latent element matrix W was generated, and the evaluation value for each product of the user who is the proposal destination was calculated using the generated latent element matrix W. In other words, the selection device 10 selects a product that is an object of electronic commerce as a proposal. However, the embodiment is not limited to this.

  For example, the selection device 10 may target any content such as news, moving images, still images, music, and microblogs. More specifically, the selection device 10 uses an evaluation value matrix R including a user's evaluation for each content to be distributed, a user feature matrix X, and a product feature matrix Y including content features, A latent element matrix W is generated, the user feature matrix X, the latent element matrix W, and the product feature matrix Y are used to estimate the evaluation of each content of the user who is the content proposal destination, and based on the estimation result Thus, the content to be proposed to the user may be selected. As described above, the selection device 10 may implement the content recommendation for the user by applying the selection process described above to the selection of the content.

[1-7. (Example of selection process)
Next, an example of the selection process executed by the selection device 10 will be described with reference to FIG. First, the selection device 10 uses the IMC to calculate a latent element matrix W based on the sparse evaluation value matrix R, the user feature matrix X, and the product feature matrix Y (step S1).

  For example, in the example illustrated in FIG. 1, the selection device 10 acquires, as the evaluation value matrix R, a matrix M10 including evaluation values for the products 1 to 5 of the users 1 to 5. For example, in the example illustrated in FIG. 1, the matrix M10 indicates that the evaluation of the “user 1” for the product B is “5” and the evaluation for the “product D” is “2”. Further, the selection device 10 acquires a matrix M20 including values indicating the characteristics of “users 1 to 5” as the user feature matrix X. For example, in the example illustrated in FIG. 1, in the matrix M <b> 20, “users 1 to 5” are associated with each row, user characteristics are associated with each column, and each user is associated with each row. Among these, “1” is stored in a column associated with the feature of the user. For example, in the example illustrated in FIG. 1, the matrix M20 indicates that “user 1” is “female” in “20's” and uses “PC (Personal Computer)” as the user terminal 100. . In the example illustrated in FIG. 1, the selection device 10 acquires, as the product feature matrix Y, a matrix M40 in which the features of the products A to E are arranged in the column direction. For example, in the example illustrated in FIG. 1, the matrix M40 indicates that “product A” belongs to “category 1” and is “free shipping”.

  When such a matrix M10, M20, M40 is acquired, the selection device 10 determines the value of each component of the matrix M30 so that the product of the matrix M20, the matrix M30, and the matrix M40 is the matrix M10. Is calculated. More specifically, the selection device 10 calculates the value of each component of the matrix M30 using a technique such as ALS. Then, the selection device 10 sets the calculated matrix M30 as the latent element matrix W. Subsequently, the selection device 10 calculates in advance a matrix M60 that becomes an integration matrix WY from the latent element matrix W and the product feature matrix Y (step S2). That is, the selection apparatus 10 calculates in advance an integration matrix WY (matrix M60) that is a product of the latent element matrix W (matrix M30) and the product feature matrix Y (matrix M40). Note that the selection device 10 can generate the restoration matrix R ′ by calculating the product of the user feature matrix X and such an integration matrix WY.

  Here, it is assumed that the selection device 10 has received a product search request from the user terminal 100 used by the user U01 (step S3). In such a case, the selection device 10 uses the WAND algorithm to specify a product that is estimated to have a high evaluation value by the user U01 (step S4). For example, the selection device 10 extracts, from the user feature matrix X, a row associated with the user U01 (for example, a row including information indicating the feature of “user 2”). In addition, the selection device 10 uses a row extracted from the user feature matrix X as a query, and uses the integration matrix WY as an index of a product (step S5). Here, considering that the product of the user feature matrix X and the integration matrix WY becomes an evaluation value matrix R (restoration matrix R ′), the integration matrix WY includes, for each feature of the user, the user's feature. It can be considered that the score of each product for the feature is included, and the value of the evaluation value can be regarded as an inner product value of the query and the score.

  Therefore, the selection device 10 uses the WAND algorithm to identify a predetermined number of products from the larger inner product value of the query and the score (step S6). For example, in the selection device 10, the inner product value of “product A” is “inner product value # 1”, and the inner product value of “product B” is “inner product value # 2” larger than “inner product value # 1”. When the inner product value of “product C” is “inner product value # 3” larger than “inner product value # 2”, “product C” and “product B” are selected. Then, the selection device 10 transmits an advertisement for a product with a high evaluation value, that is, an advertisement for “product C” and “product B”, to the user terminal 100 together with the product search result, and displays the advertisement (step S7). ). As a result, the selection device 10 can provide the user U01 with an advertisement for a product that is estimated to have a high evaluation of the user U01 using the IMC.

[2. Configuration of selection device)
Hereinafter, an example of the functional configuration of the selection device 10 that realizes the above selection process will be described. FIG. 2 is a diagram illustrating a configuration example of the selection device according to the embodiment. As illustrated in FIG. 2, the selection device 10 includes a communication unit 20, a storage unit 30, and a control unit 40.

  The communication unit 20 is realized by, for example, a NIC (Network Interface Card). The communication unit 20 is connected to the network N by wire or wireless, and transmits / receives information to / from the user terminal 100.

  The storage unit 30 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 30 also stores a product database 31, an evaluation value database 32, a user database 33, a latent element database 34, and an index database 35 (hereinafter may be collectively referred to as “each database 31 to 35”). .

  Hereinafter, an example of information registered in each of the databases 31 to 35 will be described with reference to FIGS. In the product database 31, each component of the product feature matrix Y is registered. For example, FIG. 3 is a diagram illustrating an example of information registered in the product database according to the embodiment. As shown in FIG. 3, in the product database 31, “product attributes”, which are characteristics of products targeted for electronic commerce, are associated with each row, and “product ID (Identifier)”, which is a product identifier, is stored. Associated with each column. In the product database 31, the component “1” is registered in the cell corresponding to the feature of each product.

  For example, in the example illustrated in FIG. 3, in the product database 31, among the columns associated with the product ID “A”, the product attributes “sale”, “category 2”, and “tax included” are associated with the cells. “0” is stored, and “1” is stored in the cell associated with the product attributes “category 1” and “free shipping”. This information indicates that the product indicated by the product ID “A” is a product belonging to the product attribute “category 1” and the shipping fee is free.

  In the evaluation value database 32, each component of the evaluation value matrix R is registered. For example, FIG. 4 is a diagram illustrating an example of information registered in the evaluation value database according to the embodiment. As shown in FIG. 4, in the evaluation value database 32, “user ID” for identifying each user is associated with each row, and “product ID” is associated with each column. In the evaluation value database 32, values indicating the contents of evaluation performed by each user on each product are registered in the corresponding cells.

  For example, in the example illustrated in FIG. 4, “5” is registered in the cell associated with the product ID “B” in the evaluation value database 32 in the row associated with the user ID “1”. “2” is registered in the cell associated with the product ID “D”. In such information, the user indicated by the user ID “1” assigns an evaluation value “5” to the product indicated by the product ID “B”, and assigns an evaluation value “2” to the product indicated by the product ID “D”. It shows that it did. In the evaluation value database 32, no evaluation value is registered in a cell corresponding to a product that is not evaluated by the user.

  In the user database 33, each component of the user feature matrix X is registered. For example, FIG. 5 is a diagram illustrating an example of information registered in the user database according to the embodiment. As shown in FIG. 5, in the user database 33, “user ID” for identifying each user is associated with each row, and “user” is an attribute of the user such as “male” or “female”. "Person attribute" is associated with each column. In the user database 33, the component “1” is registered in the cell corresponding to the feature of each user.

  For example, in the example shown in FIG. 5, the user attributes “female”, “20s”, and “PC” are associated with the user database 33 among the rows associated with the user ID “1”. “1” is registered in the cell, and “0” is registered in other cells. Such information indicates that the user indicated by the user ID “1” is a woman in his twenties and is using a PC as the user terminal 100.

  In the latent element database 34, values of components of the latent element matrix W generated from the user feature matrix X, the product feature matrix Y, and the evaluation value matrix R are registered. For example, FIG. 6 is a diagram illustrating an example of information registered in the latent element database according to the embodiment. As shown in FIG. 6, in the latent element database 34, “user attribute” is associated with each row, and “product attribute” is associated with each column. In the latent element database 34, values “value # 1” to “value # 11” of each component calculated from the user feature matrix X, the product feature matrix Y, and the evaluation value matrix R are registered. . In the example shown in FIG. 6, conceptual values such as “value # 1” to “value # 11” are described, but actually, the latent element database 34 includes the user feature matrix X and the product feature matrix. Numerical values calculated from Y and the evaluation value matrix R are registered. More specifically, in the latent element database 34, the product value of the user feature matrix X, the latent element matrix W, and the product feature matrix Y has values obtained so as to reproduce the evaluation value matrix R. It will be registered.

  In the index database 35, the value of each component of the integration matrix WY is registered. That is, for each user attribute, a score value indicating the relevance of each product to the user attribute is registered in the index database 35. For example, data in a format such as “User Feature ID 1: <Item ID, Score>,...” Is registered in the index database 35. Here, “User Feature ID 1” is information indicating a user attribute. “ItemID1” is a product identifier. “Score” is a score value indicating the relevance of the product indicated by the identifier of the associated product with respect to the associated user attribute.

  Returning to FIG. 2, the description will be continued. The control unit 40 is a controller. For example, various programs stored in a storage device inside the selection device 10 are stored in a RAM or the like by a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). This is realized by being executed as a work area. The control unit 40 is a controller, and may be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

  As illustrated in FIG. 2, the control unit 40 includes a request reception unit 41, an acquisition unit 42, a calculation unit 43, a search unit 44, a selection unit 45, a distribution unit 46, and a generation unit 47.

  The request receiving unit 41 receives a product search request from the user U01. More specifically, the request reception unit 41 receives a search query for a product desired by the user U01. In such a case, the request reception unit 41 searches the product database 31 for a product that matches the search query, and generates a web page indicating the search result.

  The acquisition unit 42 includes a user feature matrix X in which information indicating the characteristics of the user is arranged in the row direction, a product feature matrix Y in which information indicating the characteristics of the product to be electronic commerce is arranged in the column direction, A latent element matrix W calculated from an evaluation value matrix R in which information indicating user evaluation is arranged in the row direction, and the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y is an evaluation value. A latent element matrix W in which the value of each component is set is acquired so as to reproduce the matrix R. More specifically, the acquisition unit 42 reads out a latent element matrix W generated by a generation unit 47 described later from the latent element database 34.

  The calculation unit 43 is a value indicating the evaluation of each product of the user U01 from the product of the row including the information indicating the feature of the user U01 as the product proposal destination, the latent element matrix W, and the product feature matrix Y. Is calculated. For example, the calculation unit 43 extracts a row corresponding to the user U01 from the user feature matrix X, and calculates a product of the extracted row, the latent element matrix W, and the product feature matrix Y. That is, the calculation unit 43 estimates the evaluation of each product of the user U01 who is a product proposal destination.

  Here, when the number of products is large, each product of the user U01 is obtained from the product of the row including information indicating the characteristics of the user U01 as a product proposal destination, the latent element matrix W, and the product feature matrix Y. The calculation cost for calculating the value indicating the evaluation for the will increase. Therefore, the calculation unit 43 operates in conjunction with the search unit 44 to reduce the calculation cost for calculating a value indicating the evaluation for each product.

  For example, the calculation unit 43 calculates in advance the integration matrix WY that is the product of the latent element matrix W and the product feature matrix Y, and registers the calculated integration matrix WY in the index database 35 as the product index. On the other hand, the search unit 44 uses the information indicating the feature of the user U01 as a query, and uses the product of the latent element matrix W and the product feature matrix Y as the score of each product for the feature of the user U01. A predetermined number of products are searched from the higher inner product value calculated from the above. More specifically, the search unit 44 uses a WAND algorithm to search for a predetermined number of proposal targets from the higher inner product value calculated from the query and the score.

  As will be apparent from the description to be described later, when searching for a predetermined number of proposal targets from the higher inner product value calculated from the query and score using the WAND algorithm, the query and score to be calculated Is smaller than the inner product value held as a candidate, the calculation of the inner product value is skipped. Therefore, the search unit 44 uses the WAND algorithm only when the inner product value calculated from the query and the score is determined to be larger than the inner product value held as a candidate. The inner product value is calculated. As a result, the calculation unit 43 and the search unit 44 can reduce the calculation cost for calculating a value indicating evaluation for each product.

  The selection unit 45 selects a proposal target to be proposed to the user U01 based on the value calculated by the calculation unit 43. For example, the selection unit 45 receives notification of a predetermined number of products from the search unit 44 from the higher inner product value calculated from the query and the score. In such a case, the selection unit 45 selects a product to be actually delivered to the user U01 from the products received from the search unit 44 using a technique such as targeting.

  The distribution unit 46 distributes information for proposing the selected product to the user U01, that is, an advertisement for the selected product, to the user U01. For example, the distribution unit 46 acquires an advertisement for advertising the product selected by the selection unit 45. Then, the distribution unit 46 arranges the acquired advertisement in the web page generated by the request reception unit 41 and transmits the web page on which the advertisement is arranged to the user terminal 100. As a result, the selection device 10 can provide the user U01 with an advertisement for a product that is likely to be liked by the user U01.

  The generation unit 47 generates a latent element matrix W in which the values of the components are set so that the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y approaches the evaluation value matrix R. . For example, the generation unit 47 acquires a value indicating the user's evaluation for each product from a log server or the like, and generates an evaluation value matrix R based on the acquired value. Then, the generation unit 47 registers the generated evaluation value matrix R in the evaluation value database 32. Further, the generation unit 47 collects user features from a management server (not shown) or the like, and generates a user feature matrix X based on the collected features. Then, the generation unit 47 registers the generated user feature matrix X in the user database 33. In addition, the generation unit 47 collects product features from a management server or the like that manages the products that are subject to electronic commerce, and generates a product feature matrix Y based on the collected features. Then, the generation unit 47 registers the generated product feature matrix Y in the product database 31.

  The generation unit 47 generates a latent element matrix W from the user feature matrix X, the product feature matrix Y, and the evaluation value matrix R using an algorithm such as ALS. For example, FIG. 7 is a diagram illustrating an example of a process in which the selection device according to the embodiment calculates a latent element matrix. For example, when the algorithm such as ALS is used, it is difficult for the generation unit 47 to calculate the latent element matrix W alone. Therefore, the generation unit 47 considers the latent element matrix W to be a product of the latent element matrix U and the latent element matrix V as shown in FIG.

  For example, when the number of user features is “d1” and the number of product features is “d2”, the latent element matrix W is a d1 × d2 matrix. Therefore, the generation unit 47 regards the latent element matrix W as a product of a matrix M31 that is a d1 × k matrix, that is, a latent element matrix U, and a matrix M32 that is a k × d2 matrix, that is, a latent element matrix V. Hesitate. Here, k is an arbitrary integer. Then, the generation unit 47 uses an algorithm such as ALS so that the product of the user feature matrix X, the latent element matrix U, the latent element matrix V, and the product feature matrix Y reproduces the evaluation value matrix R. Then, the value of each component of the latent element matrix U and the latent element matrix V is calculated. Then, the generation unit 47 calculates the product of the latent element matrix U and the latent element matrix V as a latent element matrix W, and registers the calculated latent element matrix W in the latent element database 34.

[3. Example of processing executed by selection device]
Next, an example of processing for estimating an evaluation value for each product of a user when a new user is added will be described with reference to FIG. FIG. 8 is a diagram illustrating an example of processing executed when a new user is added. For example, in the example illustrated in FIG. 8, when a user indicated by the user ID “6” is added, the evaluation value matrix R includes a row corresponding to the user indicated by the user ID “6”. A row M11 in which no component value is input is added. In addition, a row M21 into which a component indicating a user attribute indicated by the user ID “6” is input is added to the user feature matrix X as a row corresponding to the user indicated by the user ID “6”. It will be.

  When the selection device 10 receives a product search request from the user indicated by the user ID “6”, the selection device 10 includes a row corresponding to the user indicated by the user ID “6” in the user feature matrix X. By calculating the product of M21, the latent element matrix W, and the product feature matrix Y, the evaluation value matrix R is input to the row M11 corresponding to the user indicated by the user ID “6”. Is calculated. And the selection apparatus 10 selects the goods proposed with respect to the user which user ID "6" shows based on the calculated value, and distributes the advertisement of the selected goods with a search result.

  In this way, the selection device 10 calculates the product of the row indicating the feature of the user who is the product proposal destination, the latent element matrix W, and the product feature matrix Y, not the entire user feature matrix X. Thus, the evaluation value for each product of the user who is the product proposal destination is estimated. For this reason, the selection device 10 has already calculated the latent element matrix W even before an evaluation value for each product by a new user is obtained, and can specify the attributes of the new user. If it is, the evaluation value with respect to each goods of the user who becomes a proposal candidate of goods can be estimated. As a result, the selection apparatus 10 can prevent an increase in calculation cost because it does not require recalculation of the latent element matrix W even when a user is added.

  Next, an example of processing for estimating an evaluation value for each product of a user when a new product is added will be described with reference to FIG. FIG. 9 is a diagram illustrating an example of processing executed when a product is added. For example, in the example illustrated in FIG. 9, when the product indicated by the product ID “F” is added, the product feature matrix Y includes a column M41 in which a component indicating the product attribute indicated by the product ID “F” is input. Will be added.

  Here, when receiving a product search request from the user indicated by the user ID “3”, the selection device 10 corresponds to the user indicated by the user ID “3” in the user feature matrix X. Row M22 is extracted. Then, the selection device 10 calculates each product of the user indicated by the user ID “3” by calculating the product of the extracted row M22, the latent element matrix W, and the product feature matrix Y to which the column M41 is added. The value of each element in the row M12 including the evaluation value for is estimated.

  Here, the latent element matrix W includes values for calculating an evaluation value for each product of the user based on the relationship between the user attribute and the product attribute. For this reason, even if a new product is registered in the product feature matrix Y, the selection device 10 can identify the attributes of the product and the existing latent elements as long as the product attribute can be specified. An evaluation value for a new product can be calculated from the matrix W and the product feature matrix Y. As a result, since the selection device 10 does not need to recalculate the latent element matrix W even when a product is added, an increase in calculation cost can be prevented.

[4. Example of WAND]
Next, an example of processing for selecting a predetermined number of products having a high user evaluation value from among products using the WAND algorithm will be described with reference to FIGS. FIG. 10 is a first diagram illustrating an example of processing for selecting a product with a high evaluation value using the WAND algorithm. FIG. 11 is a second diagram illustrating an example of processing for selecting a product having a high evaluation value using the WAND algorithm. FIG. 12 is a third diagram illustrating an example of processing for selecting a product having a high evaluation value using the WAND algorithm. In addition, in the example shown in FIGS. 10-12, it described about an example of the process which selects two goods from the one where the value of evaluation value is higher.

  For example, the selection device 10 calculates each product of the user as the proposal destination by calculating the product of the row including the information indicating the feature of the user as the proposal destination in the user feature matrix X and the integration matrix WY. The value of the evaluation value for the product can be calculated. Here, in the integration matrix WY, for each feature of the user, it can be considered that the score of each product for that feature is registered. For this reason, in the user feature matrix X, when the row including information indicating the feature of the user who is the proposal destination is regarded as a query, and the integration matrix WY is regarded as the product index, the use as the proposal destination The value of the evaluation value for each product of the person is the inner product value of the query and the score of the index of the product corresponding to the query. Then, the process of selecting a predetermined number of products from the higher evaluation value can be put into the process of performing an OR search for a document having a high score corresponding to the query.

  Therefore, the selection apparatus 10 regards a row including information indicating the feature of the user as a proposal destination in the user feature matrix X as a query, regards the integration matrix WY as a product index, and uses the WAND algorithm. A predetermined number of products are selected from those with higher evaluation values. For example, it is assumed that the user attributes “U_id_1”, “U_id_2”, and “U_id_3” are included in the features of the user as the proposal destination. In such a case, the selection device 10 extracts a combination of a product and a score associated with each of the user attributes “U_id_1”, “U_id_2”, and “U_id_3” from the integration matrix WY. In the following description, a combination of a product and a score may be described as “data”.

  For example, the selection device 10 extracts “(1, 6), (5, 4)” from the integration matrix WY as a pair of a product and a score associated with the user attribute “U_id_1”. In addition, the selection device 10 determines “(2, 3), (3, 3), (5, 1)” as a set of products and scores associated with the user attribute “U_id_2” from the integration matrix WY. To extract. In addition, the selection device 10 determines, from the integration matrix WY, “(1, 1), (2, 2), (4, 2),” as a set of products and scores associated with the user attribute “U_id_3”. (5, 1) "is extracted. Moreover, the selection apparatus 10 arranges data in order of product ID.

  Subsequently, the selection device 10 specifies the highest score (max_score) corresponding to each user attribute. For example, the selection device 10 specifies the highest score “6” corresponding to the user attribute “U_id — 1”, specifies the highest score “3” corresponding to the user attribute “U_id — 2”, and the user attribute The highest score “2” corresponding to “U_id — 3” is specified in advance. Such processing may be executed in advance when calculating the integration matrix WY.

  In addition, as illustrated in FIGS. 10A to 10C, the selection device 10 sets a pointer to the top data for each user attribute. Then, the selection device 10 identifies data including the same product ID among the data indicated by the pointer, and uses the total of the scores included in the specified data as the points of the product indicated by the product ID included in the specified data. Hold on. For example, in the example illustrated in FIG. 10, the selection device 10 includes data “(1, 6)” indicated by the pointer (A) in FIG. 10 and data (C) in FIG. 10 as data including the product ID “1”. The data “(1, 1)” indicated by the pointer is identified, and the total score “7” included in the identified data is calculated as a point. Then, as illustrated in (D) in FIG. 10, the selection device 10 holds data “(1, 7)” including the product ID “1” and the point “7” in the Accumulator.

  Next, the selection device 10 moves one pointer indicating the used data, that is, the pointers (A) and (C) in FIG. As a result, the state of each Accumulator and pointer changes to the state shown in FIG. Then, the selection device 10 repeats the same processing as that shown in FIG. For example, in the example illustrated in FIG. 11, the selection device 10 includes data “(2, 3)” indicated by the pointer (A) in FIG. 11 and data (B) in FIG. 11 as data including the product ID “2”. The data “(2, 2)” indicated by the pointer is specified, and the total score “5” included in the specified data is calculated as a point. Then, as illustrated in FIG. 11C, the selection device 10 holds data “(2, 5)” including the product ID “2” and the point “5” in the Accumulator.

  Here, referring to the point of each data registered in Accumulator, the lowest value is the point “5” of the product ID “2”. On the other hand, in consideration of the highest score corresponding to each user attribute, in order for the point of a product to exceed the value of the point “5” of the product ID “2”, data including the product ID of the product is included. , At least the user attribute “U_id_1” needs to be associated with another user attribute. For example, assuming that the data including the product ID of the product is associated with the user attribute “U_id_2” and the user attribute “U_id_3”, the maximum value of the points corresponds to the user attribute “U_id_2”. The sum of the highest score value “3” and the user attribute “U_id — 2” and the corresponding highest score value “2” is “5”, and the lowest value of the points registered in the Accumulator cannot be updated.

  Accordingly, as illustrated in FIGS. 12A and 12B, the selection device 10 uses the user attribute “U_id_2” up to the data including the product ID “5” of the data indicated by the pointer of the user attribute “U_id_1”. And the pointer of the user attribute “U_id — 3” can be advanced. That is, the selection device 10 can exclude the product ID “3” and the product ID “4” from the candidates without calculating points.

  In the example illustrated in FIG. 12, the selection device 10 calculates the data indicated by each pointer, that is, the total “6” of the scores of the product ID “5” as points. Then, as illustrated in FIG. 12C, the selection device 10 holds data “(5, 6)” including the product ID “5” and the point “6” in the Accumulator. As a result, the selection device 10 can specify the product ID “1” and the product ID “5” as the product IDs of the two products from the higher inner product value.

[5. Example of selection process flow)
Subsequently, a flow of processing in which the selection device 10 generates the latent element matrix W will be described with reference to FIG. FIG. 13 is a flowchart illustrating an example of a flow of processing in which the selection device according to the embodiment generates a latent element matrix. For example, the selection device 10 generates an evaluation value matrix R from the user's evaluation of the product (Step S101). Further, the selection device 10 generates a user feature matrix X from the user features (step S102). Further, the selection device 10 generates a product feature matrix Y from the features of the product (step S103).

  And the selection apparatus 10 produces | generates the latent element matrix W from the user feature matrix X, the goods feature matrix Y, and the evaluation value matrix R using ALS (step S104). Thereafter, the selection device 10 stores the product of the latent element matrix W and the product feature matrix Y as an index in the index database 35 (step S105), and ends the process.

  Subsequently, an example of the flow of a selection process for selecting a process proposed by the selection device 10 to the user will be described with reference to FIG. FIG. 14 is a flowchart illustrating an example of the flow of selection processing executed by the selection device according to the embodiment. First, the selection device 10 determines whether or not a request such as a search request for a product has been received from the user (step S201). If the request has not been received (step S201: No), the selection device 10 waits until it is received. When the selection device 10 accepts the request (step S201: Yes), the selection device 10 extracts the request source user row from the user feature matrix X (step S202). Subsequently, the selection device 10 uses the user attribute indicated by the extracted row as a query, considers the integration matrix WY as an index (step S203), and selects a predetermined number of products in descending order of inner product value. It specifies using WAND (step S204). The selection device 10 sets the specified product as a recommendation target for the user (step S205), proposes the recommended target product to the user together with the search result corresponding to the request (step S206), and performs processing. finish.

[6. (Modification)
In the above, an example of selection processing by the selection device 10 has been described. However, the embodiment is not limited to this. Hereinafter, the variation of the selection process which the selection apparatus 10 performs is demonstrated.

[6-1. About WAND]
The selection apparatus 10 mentioned above searched for the goods with a high value of a user's evaluation value using the WAND algorithm. However, the embodiment is not limited to this. For example, as long as the Top-k retrival algorithm is used, the selection device 10 may search for a product having a high user evaluation value using an arbitrary algorithm other than the WAND algorithm.

  In addition, the selection device 10 simply calculates the product of the row indicating the user attribute, the latent element matrix W, and the product feature matrix Y, so that the evaluation value (that is, restoration) for each product of the user. The value of the matrix R ′) may be calculated, and a predetermined number of products may be selected in descending order of the calculated evaluation value. The selection device 10 may select a product whose calculated evaluation value is higher than a predetermined threshold value, or may select a product whose calculated evaluation value falls within a predetermined range. Moreover, the selection apparatus 10 may select a predetermined number of products in descending order of the calculated evaluation value from among products not evaluated by the user, and may propose the selected products. . In addition, when the proposal target is content such as a moving image, the selection device 10 selects a predetermined number of products in descending order of the calculated evaluation value from the content not viewed by the user. The selected product may be proposed.

  Of the selection processes described above, WAND is not an essential process. That is, the selection device 10 uses the latent element matrix W generated using the IMC to estimate the evaluation value of the user with respect to the proposal target, and performs the proposal based on the estimated evaluation value as described above. The process may be executed in combination with an arbitrary process.

[6-2. (Proposed target)
In the above description, the selection device 10 proposes a product with a high evaluation value of the user calculated from the latent element matrix W to the user, with the product to be subject to electronic commerce as a proposal target. However, the embodiment is not limited to this. For example, the selection device 10 may make proposals not only for products targeted for electronic commerce but also for services.

  Further, the selection device 10 may target content such as a moving image, music, news, a web page, a user's remarks published on the web, and the like. For example, similarly to the product feature matrix Y, the selection device 10 acquires a content feature matrix Y ′ including information indicating the feature of each content. Further, the selection device 10 acquires the latent element matrix W from the user feature matrix X, the content feature matrix Y ′, and the evaluation value matrix R indicating the evaluation of the user for each content. Then, the selection device 10 estimates the user's evaluation for each content from the row including information indicating the user's characteristics, the latent element matrix W, and the content feature matrix Y ′, and based on the estimation result, You may select the content proposed to a user.

  The selection device 10 may apply the selection process described above when selecting an advertisement to be distributed to the user. For example, similarly to the product feature matrix Y, the selection device 10 acquires an advertisement feature matrix Y ″ including information indicating the feature of each advertisement. Further, the selection device 10 acquires the latent element matrix W from the user feature matrix X, the advertisement feature matrix Y ″, and the evaluation value matrix R indicating the evaluation of the user for each advertisement. Then, the selection device 10 estimates the user's evaluation for each advertisement from the row including information indicating the user's characteristics, the latent element matrix W, and the advertisement feature matrix Y ″, and based on the estimation result An advertisement to be distributed to the user may be selected.

[6-3. Device configuration〕
The selection device 10 may be connected to be able to communicate with an arbitrary number of user terminals 100. The selection device 10 may be realized by a front-end server that exchanges information with the user terminal 100 and a back-end server that selects input items. In such a case, the request receiving unit 41 and the distribution unit 46 shown in FIG. 2 are arranged in the front end server, and the acquisition unit 42, the calculation unit 43, the search unit 44, the selection unit 45, and the back end server are arranged. The generation unit 47 and the like are arranged. Further, the databases 31 to 35 stored in the storage unit 30 may be stored not in the selection device 10 but in, for example, an external storage server.

[6-4. About each database item)
In the above description, the various items included in the user feature matrix X and the product feature matrix Y are merely examples, and any items may be used as long as they are items indicating user characteristics or items indicating product characteristics. Item can be adopted.

[6-5. Others]
In addition, among the processes described in the above embodiment, all or part of the processes described as being automatically performed can be performed manually, or the processes described as being performed manually can be performed. All or a part can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the document and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in each drawing is not limited to the illustrated information.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured.

  In addition, the above-described embodiments can be appropriately combined within a range in which processing contents are not contradictory.

[7. program〕
Further, the selection device 10 according to the embodiment described above is realized by a computer 1000 having a configuration as shown in FIG. 15, for example. FIG. 15 is a diagram illustrating an example of a hardware configuration. The computer 1000 is connected to an output device 1010 and an input device 1020, and an arithmetic device 1030, a primary storage device 1040, a secondary storage device 1050, an output IF (Interface) 1060, an input IF 1070, and a network IF 1080 are connected via a bus 1090. Have

  The arithmetic device 1030 operates based on a program stored in the primary storage device 1040 and the secondary storage device 1050, a program read from the input device 1020, and the like, and executes various processes. The primary storage device 1040 is a memory device such as a RAM that temporarily stores data used by the arithmetic device 1030 for various arithmetic operations. The secondary storage device 1050 is a storage device in which data used for various calculations by the calculation device 1030 and various databases are registered, and is realized by a ROM (Read Only Memory), HDD, flash memory, or the like.

  The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various types of information such as a monitor and a printer. For example, USB (Universal Serial Bus), DVI (Digital Visual Interface), This is realized by a standard connector such as HDMI (registered trademark) (High Definition Multimedia Interface). The input IF 1070 is an interface for receiving information from various input devices 1020 such as a mouse, a keyboard, and a scanner, and is realized by, for example, a USB.

  The input device 1020 includes, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), and a PD (Phase change rewritable disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), and a tape. It may be a device that reads information from a medium, a magnetic recording medium, a semiconductor memory, or the like. The input device 1020 may be an external storage medium such as a USB memory.

  The network IF 1080 receives data from other devices via the network N and sends the data to the arithmetic device 1030, and transmits data generated by the arithmetic device 1030 to other devices via the network N.

  The arithmetic device 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic device 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040, and executes the loaded program.

  For example, when the computer 1000 functions as the selection device 10, the arithmetic device 1030 of the computer 1000 implements the function of the control unit 40 by executing a program loaded on the primary storage device 1040.

[8. effect〕
As described above, the selection device 10 includes the user feature matrix X including information indicating the features of the user, the product feature matrix Y including information indicating the features of the proposal target, and information indicating the evaluation of the user with respect to the proposal target. The component value is set such that the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y becomes the evaluation value matrix R. The obtained latent element matrix W is obtained. Then, the selection device 10 uses the information indicating the user's feature as a query, and uses the product of the latent element matrix W and the product feature matrix Y as the score of each proposal target for the user's feature, from the query and the score. A predetermined number of proposal targets are searched from the higher inner product value calculated. Then, the selection apparatus 10 selects the proposal object proposed with respect to a user based on a search result.

  As described above, the selection device 10 calculates the evaluation of the user with respect to the proposal target from the latent element matrix W calculated using the IMC. For this reason, the selection device 10 can reduce the frequency of recalculation of the latent element matrix W when a user is added, a product is added, or when a user evaluation is updated. Therefore, increase in calculation cost can be prevented. In addition, since the selection device 10 can easily calculate the evaluation of the user with respect to the proposal target from the characteristics of the user as the distribution destination, the selection target can be selected at runtime when requested by the user. Can do.

  In addition, the selection device 10 uses a Top-k retry algorithm or the like to search for a predetermined number of proposal targets from the higher inner product value calculated from the query and the score. For example, the selection device 10 uses a WAND algorithm to search for a predetermined number of proposal targets from the higher inner product value calculated from the query and the score. For this reason, the selection device 10 is estimated to have a high evaluation by the user within a realistic time even when there are a large number of proposal targets such as products that are targets of electronic commerce. The proposal target can be selected. For this reason, the selection apparatus 10 can select a proposal target at runtime.

  In addition, the selection device 10 indicates a product feature matrix Y including information indicating features of a product or service that is an object of electronic commerce as a proposal target to be proposed to the user, and shows a user's evaluation of the product or service. A latent element matrix W calculated from the evaluation value matrix R including information is acquired. In addition, the selection device 10 uses information indicating the characteristics of the user who is the proposal destination of the product or service as a query, and uses the product of the latent element matrix W and the product feature matrix Y as the product or service for the user characteristic. As a score, a predetermined number of products or services are searched from a higher inner product value calculated from the query and the score. And the selection apparatus 10 selects the goods or service proposed with respect to a user based on a search result. For this reason, the selection apparatus 10 can perform selection of the goods and service used as proposal object at runtime.

  The selection device 10 also includes a user feature matrix X in which information indicating user characteristics is arranged in the row direction, a product feature matrix Y in which information indicating features of the proposal target is arranged in the column direction, and a user for the proposal target. Is a latent element matrix W calculated from an evaluation value matrix R in which information indicating the evaluation is arranged in the row direction, and the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y is the evaluation value matrix R The latent element matrix W in which the component values are set so as to be obtained is acquired. Further, the selection device 10 determines the user who is the proposal target of the proposal target from the product of the row including information indicating the characteristics of the user who is the proposal target of the proposal target, the latent element matrix W, and the product feature matrix Y. The value which shows evaluation with respect to each proposal object of is calculated. And the selection apparatus 10 selects the proposal object proposed with respect to a user based on the calculated value.

  As described above, the selection device 10 calculates the evaluation of the user with respect to the proposal target from the latent element matrix W calculated using the IMC. For this reason, the selection apparatus 10 becomes the proposal target proposal destination only by calculating the product of the row indicating the attribute of the user who is the proposal target of the proposal target, the latent element matrix W, and the product feature matrix Y. It is possible to easily obtain an evaluation for each proposal target of the user.

  Further, the selection device 10 extracts a row including information indicating the feature of the user who is the proposal target of the proposal target from the user feature matrix X, and the extracted row, the latent element matrix W, and the product feature matrix Y From the product, a value indicating the evaluation of each proposal target by the user who is the proposal target of the proposal target is calculated. For this reason, the selection apparatus 10 can prevent an increase in calculation cost when calculating an evaluation for each proposal target of a user who is a proposal target of the proposal target.

  In addition, the selection device 10 includes, in the row direction, a product feature matrix Y in which information indicating the characteristics of the product or service that is the object of the electronic commerce is arranged in the column direction, and information indicating the evaluation of each user for the product or service in the row direction. A latent element matrix W calculated from the arranged evaluation value matrix R is acquired. In addition, the selection device 10 evaluates the user's product or service from the product of the row including the information indicating the user's feature that is the proposal destination of the product or service, the latent element matrix W, and the product feature matrix Y. A value indicating is calculated. And the selection apparatus 10 selects the goods or service proposed with respect to a user based on a calculation result. For this reason, the selection apparatus 10 can perform selection of the goods and service used as proposal object at runtime.

  The selection device 10 also includes a user feature matrix X including information indicating user characteristics, a product feature matrix Y including information indicating content characteristics, and an evaluation value matrix including information indicating user evaluation of the contents. A latent element matrix W calculated from R, in which the component values are set such that the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y becomes the evaluation value matrix R Get W. Further, the selection device 10 evaluates the content of the user who is the content proposal destination using the information indicating the characteristics of the user who is the content proposal destination, the latent element matrix W, and the product feature matrix Y. presume. And the selection apparatus 10 selects the content proposed with respect to a user based on an estimation result.

  As described above, the selection device 10 calculates the evaluation of the user for the content to be distributed from the latent element matrix W calculated using the IMC, and selects the content to be proposed according to the calculation result. . For this reason, the selection device 10 can reduce the frequency of recalculation of the latent element matrix W when a user is added, a product is added, or when a user evaluation is updated. Therefore, increase in calculation cost can be prevented.

  Further, the selection device 10 generates a latent element matrix W in which the values of the respective components are set so that the product of the user feature matrix X, the latent element matrix W, and the product feature matrix Y approaches the evaluation value matrix R. More specifically, the selection device 10 includes the latent element matrix U having the same number of rows as the number of columns of the user feature matrix X and the latent element matrix V having the same number of columns as the number of rows of the product feature matrix Y, The latent element matrix U and the latent element matrix V are generated so that the product of the feature matrix X, the latent element matrix U, the latent element matrix V, and the product feature matrix Y reproduces the evaluation value matrix R. Then, the selection device 10 sets the product of the generated latent element matrix U and the latent element matrix V as a latent element matrix W. For this reason, the selection apparatus 10 can generate the latent element matrix W using a technique such as ALS, for example.

  As described above, some of the embodiments of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various modifications, including the aspects described in the disclosure section of the invention, based on the knowledge of those skilled in the art, It is possible to implement the present invention in other forms with improvements.

  Moreover, the above-mentioned “section (module, unit)” can be read as “means”, “circuit”, and the like. For example, the distribution unit can be read as distribution means or a distribution circuit.

DESCRIPTION OF SYMBOLS 10 Selection apparatus 20 Communication part 30 Storage part 31 Goods database 32 Evaluation value database 33 User database 34 Latent element database 35 Index database 40 Control part 41 Request reception part 42 Acquisition part 43 Calculation part 44 Search part 45 Selection part 46 Distribution part 47 Generation | occurrence | production Department 100 User terminal

Claims (15)

Calculated from a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the proposal target, and a third matrix including information indicating the evaluation of the user with respect to the proposal target An acquisition unit for acquiring a fourth matrix, wherein a value of a component is set such that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix;
The information indicating the feature of the user who is the proposal target of the proposal target is used as a query, and the value of the product of the fourth matrix and the second matrix is used as the score of each proposal target with respect to the feature of the user. A search unit for searching for a predetermined number of proposal targets from a higher inner product value calculated from the score;
A selection unit that selects a proposal target to be proposed to the user based on a search result by the search unit; The selection according to claim 1, wherein the search unit searches for a predetermined number of proposal targets from a higher inner product value calculated from the query and the score using a WAND algorithm. apparatus. The acquisition unit indicates a second matrix including information indicating characteristics of a product or service that is an object of electronic commerce as a proposal target to be proposed to the user, and an evaluation of the user with respect to the product or service Obtaining a fourth matrix calculated from the third matrix including information;
The search unit uses, as a query, information indicating a feature of a user who is a proposal destination of the product or service, and uses the product of the fourth matrix and the second matrix as the product or service for the user feature. As a score of, search for a predetermined number of the products or services from the higher inner product value calculated from the query and the score,
The selection device according to claim 1, wherein the selection unit selects a product or service to be proposed to the user based on a search result by the search unit. A first matrix in which information indicating the features of the user is arranged in the row direction, a second matrix in which information indicating the features of the proposal target is arranged in the column direction, and information indicating the evaluation of the user with respect to the proposal target in the row direction The fourth matrix calculated from the arranged third matrix, and the value of the component is set so that the product of the first matrix, the fourth matrix, and the second matrix becomes the third matrix An acquisition unit for acquiring a fourth matrix;
For each proposal target of the user who is the proposal target of the proposal target from the product of the row including the information indicating the characteristics of the user who is the proposal target of the proposal target, the fourth matrix, and the second matrix A calculation unit for calculating a value indicating evaluation;
And a selection unit that selects a proposal target to be proposed to the user based on the value calculated by the calculation unit. The calculation unit extracts a row including information indicating characteristics of a user who is a proposal destination of the proposal target from the first matrix, and a product of the extracted row, the fourth matrix, and the second matrix The value which shows the evaluation with respect to each proposal object of the user who becomes a proposal destination of the proposal object is calculated from. The selection device according to claim 4 characterized by things. The acquisition unit, as a proposal target to be proposed to the user, a second matrix in which information indicating characteristics of a product or service that is an object of electronic commerce is arranged in a column direction, and each use for the product or service A fourth matrix calculated from the third matrix in which information indicating the evaluation of the person is arranged in the row direction,
The calculation unit calculates the product or service for the user from the product of the row including information indicating the characteristics of the user to whom the product or service is proposed, the fourth matrix, and the second matrix. Calculate a value indicating the evaluation,
The selection device according to claim 4 or 5, wherein the selection unit selects a product or service to be proposed to the user based on a calculation result by the calculation unit. A fourth matrix calculated from a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the content, and a third matrix including information indicating the evaluation of the user with respect to the content. An acquisition unit that acquires a fourth matrix, the value of which is set so that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix;
An estimation unit that estimates an evaluation of the content of the user who is the content proposal destination by using information indicating characteristics of the user who is the content proposal destination, the fourth matrix, and the second matrix When,
And a selection unit that selects content to be proposed to the user based on a result of estimation by the estimation unit. A generating unit configured to generate the fourth matrix in which values of each component are set so that a product of the first matrix, the fourth matrix, and the second matrix approaches the third matrix; The selection device according to any one of claims 1 to 7. The generation unit includes a fifth matrix having the same number of rows as the number of columns of the first matrix, and a sixth matrix having a number of columns equal to the number of rows of the second matrix, wherein the first matrix and the fifth matrix And the sixth matrix and the second matrix generate the fifth matrix and the sixth matrix so that the product of the third matrix is reproduced, and the product of the generated fifth matrix and sixth matrix is the fourth matrix. The selection device according to claim 8, wherein: A selection method performed by a selection device, comprising:
Calculated from a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the proposal target, and a third matrix including information indicating the evaluation of the user with respect to the proposal target An acquisition step of acquiring a fourth matrix, wherein a value of a component is set such that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix;
The information indicating the feature of the user who is the proposal target of the proposal target is used as a query, and the value of the product of the fourth matrix and the second matrix is used as the score of each proposal target with respect to the feature of the user. A search step of searching for a predetermined number of proposal targets from the higher inner product value calculated from the score;
And a selection step of selecting a proposal target to be proposed to the user based on a search result obtained by the search step. A selection method performed by a selection device, comprising:
A first matrix in which information indicating the features of the user is arranged in the row direction, a second matrix in which information indicating the features of the proposal target is arranged in the column direction, and information indicating the evaluation of the user with respect to the proposal target in the row direction The fourth matrix calculated from the arranged third matrix, and the value of the component is set so that the product of the first matrix, the fourth matrix, and the second matrix becomes the third matrix An acquisition step of acquiring a fourth matrix;
For each proposal target of the user who is the proposal target of the proposal target from the product of the row including the information indicating the characteristics of the user who is the proposal target of the proposal target, the fourth matrix, and the second matrix A calculation step of calculating a value indicating evaluation;
And a selection step of selecting a proposal target to be proposed to the user based on the value calculated by the calculation step. A selection method performed by a selection device, comprising:
A fourth matrix calculated from a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the content, and a third matrix including information indicating the evaluation of the user with respect to the content. An acquisition step of acquiring a fourth matrix, the value of which is set so that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix;
An estimation step of estimating evaluation of the content of the user who is the content proposal destination using information indicating the characteristics of the user who is the content proposal destination, the fourth matrix, and the second matrix When,
And a selection step of selecting content to be proposed to the user based on the estimation result of the estimation step. Calculated from a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the proposal target, and a third matrix including information indicating the evaluation of the user with respect to the proposal target An acquisition procedure for acquiring a fourth matrix, wherein a value of a component is set such that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix;
The information indicating the feature of the user who is the proposal target of the proposal target is used as a query, and the value of the product of the fourth matrix and the second matrix is used as the score of each proposal target with respect to the feature of the user. A search procedure for searching for a predetermined number of proposal targets from the higher inner product value calculated from the score,
A selection program for causing a computer to execute a selection procedure for selecting a proposal target to be proposed to the user based on a search result obtained by the search procedure. A first matrix in which information indicating the features of the user is arranged in the row direction, a second matrix in which information indicating the features of the proposal target is arranged in the column direction, and information indicating the evaluation of the user with respect to the proposal target in the row direction The fourth matrix calculated from the arranged third matrix, and the value of the component is set so that the product of the first matrix, the fourth matrix, and the second matrix becomes the third matrix An acquisition procedure for acquiring a fourth matrix;
For each proposal target of the user who is the proposal target of the proposal target from the product of the row including the information indicating the characteristics of the user who is the proposal target of the proposal target, the fourth matrix, and the second matrix A calculation procedure for calculating a value indicating evaluation;
A selection program for causing a computer to execute a selection procedure for selecting a proposal target to be proposed to the user based on a value calculated by the calculation procedure. A fourth matrix calculated from a first matrix including information indicating the characteristics of the user, a second matrix including information indicating the characteristics of the content, and a third matrix including information indicating the evaluation of the user with respect to the content. An acquisition procedure for acquiring a fourth matrix, the component values of which are set such that a product of the first matrix, the fourth matrix, and the second matrix is the third matrix;
An estimation procedure for estimating the evaluation of the content of the user who is the content proposal destination by using the information indicating the characteristics of the user who is the content proposal destination, the fourth matrix, and the second matrix When,
A selection program for causing a computer to execute a selection procedure for selecting content to be proposed to the user based on an estimation result obtained by the estimation procedure.

Images