WO2023085165A1 - Dispositif de recommandation d'article et procédé de recommandation d'article - Google Patents

Dispositif de recommandation d'article et procédé de recommandation d'article Download PDF

Info

Publication number
WO2023085165A1
WO2023085165A1 PCT/JP2022/040787 JP2022040787W WO2023085165A1 WO 2023085165 A1 WO2023085165 A1 WO 2023085165A1 JP 2022040787 W JP2022040787 W JP 2022040787W WO 2023085165 A1 WO2023085165 A1 WO 2023085165A1
Authority
WO
WIPO (PCT)
Prior art keywords
customer
product
subject
preferences
model
Prior art date
Application number
PCT/JP2022/040787
Other languages
English (en)
Japanese (ja)
Inventor
淳一 近添
チュン クアン ファム
Original Assignee
株式会社アラヤ
大学共同利用機関法人自然科学研究機構
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 株式会社アラヤ, 大学共同利用機関法人自然科学研究機構 filed Critical 株式会社アラヤ
Priority to JP2023559578A priority Critical patent/JPWO2023085165A1/ja
Publication of WO2023085165A1 publication Critical patent/WO2023085165A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]

Definitions

  • the present invention relates to an article recommendation device and an article recommendation method. More specifically, the present invention relates to an item recommendation device that uses collaborative filtering to select the item most suitable for a customer's taste and/or purchasing tendency from a large number of items, and an item recommendation method for executing this.
  • the “goods” of the product recommendation device and the product recommendation method are not limited to tangible products, but include intangible products such as software, and services such as online games. In other words, regardless of whether it is tangible or intangible, it refers to "things" that are subject to commercial transactions.
  • E-commerce electronic commerce
  • An EC site which is a website that implements EC, recommends products according to individual tastes for the purpose of encouraging consumer purchasing behavior.
  • a method of accumulating purchase data of other consumers and making inferences using information of other consumers with similar tastes is widely used. This reasoning is called collaborative filtering.
  • Patent Document 1 describes a product recommendation system that presents appropriate products to be recommended for existing customers when replacing products, such as durable consumer goods that have a long replacement cycle and products for corporations that emphasize practicality rather than taste.
  • a product recommendation method and a recording medium recording the product recommendation method are disclosed.
  • Documents describing collaborative filtering are disclosed in Non-Patent Document 1 and Non-Patent Document 2.
  • Documents describing independent component analysis are disclosed in Non-Patent Document 3 and Non-Patent Document 4.
  • the present invention solves these problems, and is capable of selecting and recommending products that match the customer's tastes and/or purchasing tendencies with higher estimation accuracy than conventional methods. intended to provide
  • the item recommendation apparatus of the present invention includes a neural network arithmetic processing unit that imitates the preferences of a plurality of subjects and outputs an estimated value of the subjects' preferences with respect to input image data of an item. , a customer preference model formation processing unit that generates a customer preference model, which is a functional model for estimating a customer's preference from an estimated value of a subject's preference, in accordance with a predetermined calculation procedure; and a collaborative filtering arithmetic processing unit that reads the customer's preference for a plurality of items, calculates the degree of recommendation of the item to the customer, and selects the item to be recommended to the customer based on the degree of recommendation. do.
  • FIG. 1 is a schematic diagram showing an example of an article recommendation device according to an embodiment of the present invention and its usage pattern;
  • FIG. It is a block diagram which shows the hardware constitutions of a server.
  • 3 is a block diagram showing the hardware configuration of a client;
  • FIG. FIG. 3 is a block diagram showing the software functions of the client;
  • 2 is a block diagram showing software functions of the server;
  • FIG. It is a figure which shows the field structure of various tables.
  • 4 is a procedure manual (flowchart) showing the flow of work in the server;
  • FIG. 4 is a schematic diagram showing an outline of the operation in the learning mode of the NN arithmetic processing unit;
  • FIG. 4 is a schematic diagram showing an overview of the operation in the estimation mode of the NN arithmetic processing unit;
  • FIG. 10 is a diagram schematically showing how values in each table are entered according to the work flow in the server;
  • FIG. 4 is a schematic diagram for explaining the content of processing on a classification basis of the article recommendation device according to the embodiment of the present invention, including comparison with the conventional technology;
  • FIG. 4 is a diagram for explaining the mechanisms of correlation coefficients and collaborative filtering in the prior art;
  • 10 is a graph showing estimated results with respect to the data loss rate when the estimated results of the item recommendation device according to the embodiment of the present invention are confirmed by simulation.
  • Fig. 10 is a graph showing the estimated performance of regular collaborative filtering and collaborative filtering with an AI agent using the Moußs 100K dataset.
  • FIG. 11 is a block diagram showing software functions of a server according to a modification of the present invention. It is a figure which shows the field structure of various tables. 4 is a procedure manual (flowchart) showing the flow of work in the server; Schematic diagram A showing an overview of the operation of the function model calculation processing unit when forming a function model that imitates customer preferences, and an overview of the operation when estimating the customer's preference for the target product. Schematic diagram B. FIG.
  • a product recommendation apparatus uses an artificial neural circuit that performs image recognition to create an AI agent that imitates preferences from the preference data of tens to hundreds of subjects for each subject.
  • an AI agent that imitates preferences from the preference data of tens to hundreds of subjects for each subject.
  • the AI agent created in this way can estimate preferences with sufficient accuracy even for new samples, so even for products that have not yet been released, preference data for tens to hundreds of people can be collected. can be generated.
  • collaborative filtering By applying collaborative filtering using newly generated preference data, it is possible to make appropriate recommendations to desired customers even for completely new products.
  • subjects a plurality of persons who provide learning data for the learning algorithm
  • a person for whom the degree of recommendation is calculated will be referred to as a customer.
  • customers can also be subjects, so in order to avoid confusion, we will call them separately.
  • FIG. 1 is a schematic diagram showing an example of an item recommendation device according to an embodiment of the present invention and its usage pattern.
  • the server 101 constituting the article recommendation device includes a web server function, a DNN (Deep Neural Network) function, and a function of calculating the degree of product recommendation based on collaborative filtering.
  • a server 101 and clients 102 a and 102 b are connected by a well-known network 103 .
  • the notebook PC placed above the server 101 indicates the client 102a in the learning mode.
  • the client 102a is caused to read the product image data from the USB memory 104 or the like, and the test subject's answers to the questionnaire (not shown) are input.
  • a notebook PC placed below the server 101 indicates the client 102b in the estimation mode.
  • the client 102b conducts questionnaires regarding the preferences of several to several tens of products to target customers, and saves the results. Based on the result, the client 102b calculates the estimation result of the subject AI and the correlation coefficient, and finally calculates the degree of recommendation of the product to the customer. The calculated recommendation degree is displayed on the display screen V105 of the client 102b.
  • the clients 102a and 102b are referred to as the client 102 when not distinguished from each other.
  • a client-server format which is probably the most realistic form of use, is shown as an example of a device that estimates and calculates the degree of recommendation of goods, services, etc.
  • the article recommendation device according to the present invention can also be implemented in a stand-alone configuration.
  • the server 101 can also be configured in the cloud.
  • FIG. 2 is a block diagram showing the hardware configuration of the server 101.
  • a CPU 201 a ROM 202 , a RAM 203 , a nonvolatile storage 204 and a NIC (Network Interface Card) 205 are connected to a bus 208 .
  • a well-known PC or the like can be substituted for the server 101, in which case a display unit 206 and an operation unit 207 are attached.
  • FIG. 3 is a block diagram showing the hardware configuration of the client 102.
  • a client 102 such as a well-known PC or the like has a CPU 301 , a ROM 302 , a RAM 303 , a nonvolatile storage 304 , a NIC 305 , a display section 306 and an operation section 307 connected to a bus 308 .
  • the hardware configuration of the client 102 is substantially the same as that of the server 101 except that it has a display unit 306 and an operation unit 307 .
  • FIG. 4 is a block diagram illustrating the software functionality of client 102 .
  • the client 102 uses the input/output control unit 401, which is a function of a web browser or the like, to transmit operation instructions issued from the operation unit 307 to the server 101, receive execution results from the server 101, and display them on the display unit 306. .
  • FIG. 5 is a block diagram showing software functions of the server 101.
  • the input/output control unit 501 receives operation commands and predetermined data from the client 102 , exchanges data with various tables and arithmetic processing functions, and transmits various execution results to the client 102 .
  • the data structures of various tables will be described later in FIG.
  • a table 509, a target user correlation coefficient table 510, and a target user collaborative filtering table 511 are listed.
  • a neural network arithmetic processing unit (hereinafter “NN arithmetic processing unit”) 502 is an engine of the DNN.
  • the NN arithmetic processing unit 502 estimates and outputs preference data that mimics the preferences of a predetermined subject.
  • Collaborative filtering arithmetic processing unit 503 includes correlation coefficient arithmetic processing unit 504 and executes known collaborative filtering arithmetic processing.
  • FIG. 6 is a diagram showing field configurations of various tables shown in the block diagram of FIG.
  • the product master 505 has a product ID field and a product image data field.
  • a product ID that uniquely identifies a product is stored in the product ID field.
  • the product image data field stores product image data. Therefore, the product image data field is a variable length field.
  • the file name of the product image file may be stored instead of the product image data field.
  • the AI user coefficient table 506 has a subject ID field and an approximation function parameter field.
  • a subject ID that uniquely identifies a subject is stored in the subject ID field.
  • the approximation function parameter field stores the approximation function parameter configuring the subject's AI.
  • the AI user answer result table 507 has a subject ID field, a product ID field, and a favorability field.
  • the subject ID field is the same as the same name field of the AI User Coefficient Table 506 .
  • the product ID field is the same as the same name field of the product master 505 .
  • the likability field stores the likability value answered by the subject to the product associated with the product ID stored in the product ID field.
  • the AI user estimation result table 508 has a subject ID field, a product ID field, and a favorability field.
  • the field structure of the AI user estimation result table 508 is the same as the field structure of the AI user answer result table 507 .
  • the likability field stores the likability value answered by the subject's AI, not the subject's answer, for the product associated with the product ID stored in the product ID field.
  • the target user answer result table 509 has a customer ID field, a product ID field, and a favorability field.
  • a customer ID that uniquely identifies a target user who is a customer is stored in the customer ID field.
  • the product ID field and favorability field are the same as the same name field of the AI user answer result table 507 .
  • the target user correlation coefficient table 510 has a customer ID field, a subject ID field, a correlation coefficient field and a selection flag field.
  • the customer ID field is the same as the same name field of the target user answer result table 509 .
  • the subject ID field is the same as the field of the same name in the AI user answer result table 507 and the like.
  • the correlation coefficient field stores the correlation coefficient between the customer identified in the customer ID field and the subject identified in the subject ID field. This correlation coefficient is calculated by calculating the correlation coefficient between the favorability in the AI user answer result table 507 and/or the AI user estimation result table 508 and the favorability in the target user answer result table 509 for the same product ID. be done.
  • the selection flag field stores a selection flag that identifies an AI subject and is referred to when calculating the degree of product recommendation for a customer identified by a customer ID. for example, - For a subject ID with a correlation coefficient of 0.5 or more, - A selection flag is attached to the subject IDs of the top 20 subjects under conditions such as setting the logic of the selection flag to "true".
  • the target user collaborative filtering table 511 has a customer ID field, a product ID field, and a product recommendation degree field.
  • the customer ID field and product ID field are the same as the fields with the same names in the target user answer result table 509 and the like.
  • the product recommendation level field stores a product recommendation level indicating how much the product specified by the product ID in the product ID field can be recommended to the customer. Specifically, the favorability of the product ID specified by the product ID in the product ID field of the AI user estimation result table 508 is multiplied by the correlation coefficient of the target user correlation coefficient table 510, and the sum of these values is summed. The value divided by the number of elements is stored.
  • FIG. 7 is a flow chart showing the work flow (procedure) of the article recommendation method in the server 101 .
  • an AI agent is built in the server 101 based on the questionnaire results of a plurality of subjects.
  • the operator of the server 101 shows tens to hundreds of product images to tens to hundreds of test subjects, and asks them to answer the positive impressions of the product images. For example, the subject is asked to respond with a five-level evaluation such as "very much/somewhat like/neither/somewhat dislike/very dislike". Then, the operator stores the content of the subject's answer in the AI user answer result table 507 (S702).
  • the operator uses the content stored in the AI user answer result table 507 in step S702 and the product image data stored in the product master 505 to cause the DNN to perform learning, thereby creating an AI agent for each subject.
  • the generated approximation function parameters are stored in the AI user coefficient table 506 (S703).
  • the steps S702 and S703 described above are at least work to be prepared in advance prior to the operation of product recommendation in the server 101 .
  • FIG. 8 is a schematic diagram showing an overview of the operation of the NN arithmetic processing unit 502 in the learning mode.
  • the NN arithmetic processing unit 502 designates a learning mode, for example, loads image data of clothes as learning data, and labels (answers) the values of five-level evaluation as preference data when subject A sees clothes. to enter. Then, the NN arithmetic processing unit 502 generates the subject A's approximate function parameter P801a. Then, the NN arithmetic processing unit 502 updates the approximate function parameters of the subject A by performing this learning process on several tens to several hundreds of pieces of image data of different clothes.
  • the NN arithmetic processing unit 502 performs similar processing on subject B, subject C, . . . subject n.
  • the approximate function parameters P801a, P801b . . . P801n for each subject thus created are stored in the AI user coefficient table 506. This is the process of step S703.
  • This step S703 completes an AI agent that imitates the subject's preferences. By reading unknown product image data, the AI agent can estimate how favorable the test subject is.
  • FIG. 9 is a schematic diagram showing an overview of the operation of the NN arithmetic processing unit 502 in the estimation mode.
  • the NN arithmetic processing unit 502 designates the estimation mode, incorporates the approximate function parameter P801a of the subject A, and then reads the image data of the unknown clothes as learning data. Then, the NN arithmetic processing unit 502 becomes an AI agent of the subject A, and estimates preference data that imitates the preference when the subject A sees image data of unknown clothes.
  • the NN arithmetic processing unit 502 performs this estimation process to estimate several to several tens of image data of the clothes answered in the questionnaire by the customer X, thereby determining the subject A's preference for the clothes answered in the questionnaire by the customer X.
  • the NN arithmetic processing unit 502 performs similar processing on subject B, subject C, . . . subject n.
  • the estimated preference data for each subject thus generated is stored in the AI user estimation result table 508 . This process is the process of step S704, which will be described later.
  • the operator of the server 101 conducts a questionnaire on favorable impressions of several to several tens of products to the customer who presents the degree of product recommendation. Then, the operator stores the contents of the responses to the questionnaire in the target user response result table 509 (S704).
  • the input/output control unit 501 through the NN arithmetic processing unit 502, estimates the favorability of several to several tens of products answered by the customer with the AI agent of the plurality of subjects created in step S703, and AI It is stored in the user estimation result table 508 (S705).
  • the input/output control unit 501 compares the target user answer result table 509 and the AI user estimation result table 508 through the correlation coefficient calculation processing unit 504 to calculate the correlation coefficient between the customer and the subject.
  • the input/output control unit 501 then stores the correlation coefficient in the target user correlation coefficient table 510 (S706).
  • the input/output control unit 501 - For a subject ID with a correlation coefficient of 0.5 or more, Under a condition such as targeting the subject IDs of the top 20 subjects, logical "true" is added to the selection flag field of the subject in the target user correlation coefficient table 510 and stored (S707).
  • the correlation coefficient between the customer and the subjects can be immediately calculated.
  • sites where data processing that employs collaborative filtering, such as e-commerce sites it is often the case that correlation coefficients cannot be calculated. This is because the product image used in the customer's questionnaire is not used in the subject's questionnaire, and therefore may be an unknown product image for the subject.
  • the subject's AI agent since the subject's AI agent is completed, if the product image used in the customer's questionnaire is read, the subject's AI agent will estimate the degree of favorability that the subject shows. and based on that estimate, a correlation coefficient can be calculated.
  • the input/output control unit 501 uses the AI agent of the subject selected in step S707 to estimate, through the NN arithmetic processing unit 502, the favorability of the candidate product to be recommended to the customer. It is stored in the estimation result table 508 (S708). Finally, the input/output control unit 501, through the collaborative filtering arithmetic processing unit 503, calculates the degree of recommendation of the product to be recommended to the customer based on the correlation coefficient calculated in step S706 and the test subject's favorable rating estimated in step S708. do. Then, the input/output control unit 501 stores the calculated recommendation level of the product recommended to the customer in the product recommendation level field of the target user collaborative filtering table 511 (S709), and ends the series of processes (S710).
  • FIG. 10 is a diagram schematically showing how values are entered in each table according to the product recommendation work flow in the server 101 .
  • step S702 subjects A to E look at several tens to several hundreds of product images of learning products and answer their favourability.
  • the reply data T1001 which is the content of this reply, is stored in the AI user reply result table 507 by the server 101.
  • FIG. The server 101 reads the answer data T1001 into the NN arithmetic processing unit 502 in step S703, and forms an AI agent for each subject.
  • the approximation function parameters created at this time are stored in the AI user coefficient table 506 .
  • step S704 the operator of the server 101 conducts a questionnaire on the favorability of several to several tens of correlation calculation products to the customer X who presents the degree of product recommendation.
  • Customer response data T1003 which is the contents of the questionnaire response, is stored in the target user response result table 509. FIG.
  • the correlation coefficient can be calculated. Therefore, in step S706, the input/output control unit 501 compares the target user answer result table 509 and the AI user estimation result table 508 through the NN calculation processing unit 502 to calculate the correlation coefficient.
  • the obtained correlation coefficient T1005 is stored in the target user correlation coefficient table 510.
  • FIG. based on the correlation coefficient T1005 obtained by the calculation in step S706, it is possible to identify the subject AI whose recommendation degree is to be calculated based on a predetermined selection criterion.
  • the selection flag field of the subject is assigned a logical “true”. This is the selection flag T1006.
  • step S708 the input/output control unit 501 causes the test subject AI whose selection flag is logically “true” to read the product image data of the recommended estimation calculation target product through the NN calculation processing unit 502, and estimating the favorability of The estimation data T1007, which is the estimation result, is stored in the AI user estimation result table 508.
  • FIG. 1 the input/output control unit 501 causes the test subject AI whose selection flag is logically “true” to read the product image data of the recommended estimation calculation target product through the NN calculation processing unit 502, and estimating the favorability of The estimation data T1007, which is the estimation result, is stored in the AI user estimation result table 508.
  • the degree of recommendation for the customer X is calculated in step S709.
  • the recommendation level data T1008, which is the calculation result, is stored in the product recommendation level field of the target user collaborative filtering table 511.
  • the operation of the article recommendation apparatus described above with reference to FIGS. 7 to 10 has been described on the premise that the input/output control unit 501 executes regression in the learning algorithm through the NN arithmetic processing unit 502 that executes DNN. i.e. ⁇ 1>
  • the input/output control unit 501 creates an AI agent that imitates the subject's preference based on a questionnaire given to the subject and outputs a preference score for an arbitrary product.
  • the input/output control unit 501 calculates the correlation between the output score of the AI agent and the target customer's preference score.
  • the input/output control unit 501 recommends products based on this correlation score. This is the process.
  • the experiment to be described later with reference to FIG. 13 also has five levels of responses to the questionnaire, and is performed by regression processing.
  • the input/output control unit 501 creates an AI agent that outputs purchase behavior reading predictions for arbitrary products based on the customer's purchase behavior data.
  • the input/output control unit 501 calculates the correlation with the purchase behavior of the target customer based on the output score of the AI agent.
  • the input/output control unit 501 recommends products based on this correlation score. Classification-based processing is also possible. Purchasing behavior such as purchasing a product or not purchasing a recommended product corresponds to classification in a learning algorithm.
  • FIG. 11A and 11B are schematic diagrams for explaining the details of the classification-based processing of the article recommendation device according to the embodiment of the present invention, including a comparison with the conventional technology.
  • FIG. 12 is a diagram for explaining the mechanisms of correlation coefficients and collaborative filtering in the prior art.
  • the collaborative filtering arithmetic processing unit of the conventional technology includes the purchase behavior data of consumers A, B, C, D, . . . corresponding to subjects in the embodiment of the present invention, A correlation coefficient is calculated by comparing with purchasing behavior data. Then, the collaborative filtering arithmetic processing unit of the conventional technology recommends the optimal product to the consumer X based on the purchasing behavior data of the consumers who have a high correlation coefficient with the consumer X, that is, the consumers who have similar tastes. .
  • the row labeled “recommendation degree” is the computation result of collaborative filtering of the product for consumer X.
  • FIG. 12 the row labeled “recommendation degree” is the computation result of collaborative filtering of
  • the article recommendation device makes an AI agent learn the preferences of each consumer based on the purchase behavior data of consumers A, B, C, D, . . . . Then, for consumer A, an AI agent that has learned consumer A's preferences and imitates consumer A's preferences is created. Similarly, an AI agent that imitates consumer B's preferences, an AI agent that imitates consumer C's preferences, an AI agent that imitates consumer D's preferences, and so on are created.
  • the article recommendation device outputs the output of an AI agent having a high correlation coefficient with consumer X, i.e., an AI agent whose taste is very similar to that of consumer X (for example, the taste of consumer B is imitated).
  • the optimal product is recommended to the consumer X from the virtual purchasing behavior data of the AI agent).
  • FIG. 12 is a diagram for explaining a purchase history on an EC site.
  • the purchase history of many e-commerce sites includes information such as "a customer bought a certain product” and "a customer did not buy a certain product even though it was recommended”. There is information such as "I don't know if I will buy a certain product or not.” In other words, the status is that consumers' purchasing behavior is unknown.
  • the column marked with "1" at the intersection of "merchandise” and “consumer” corresponds to the information "a certain customer bought a certain merchandise”.
  • a column in which "0" is written at the intersection of "product” and “consumer” corresponds to information that "a certain customer did not buy a certain product even though it was recommended.”
  • the column marked with "-" at the intersection of "product” and “consumer” corresponds to the information that "a certain customer does not know whether or not to buy a certain product”.
  • a large-scale EC site has millions to tens of millions of products. It is nearly impossible for a single customer to decide whether or not to purchase all of the products. Therefore, the purchase history includes information such as "I don't know” in addition to "I bought” and "I didn't buy”.
  • a group of commodities for which consumer A's purchasing behavior is clear is defined as a consumer A commodities group.
  • a group of commodities for which consumer B's purchasing behavior is clear is defined as a consumer B commodities group.
  • a group of products for which consumer X's purchasing behavior is clear is defined as a consumer X product group.
  • Product purchasing tendencies vary widely due to the personal circumstances of individual consumers. Therefore, among consumers A, B, C, D, and so on, the possibility that there is a consumer who has a group of products that completely match consumer X's product group is determined by the type of product purchased by consumer X. In other words, the more product groups there are, the lower the price.
  • the consumers who match the consumer X product group are the consumers A and C, and the consumer B, consumer D, and consumer E have unknown purchasing behavior for the consumer X product group. there is a product.
  • Calculation of the correlation coefficient in the conventional technology has been performed for a group of products that match each other, out of the consumer X product group and the consumer B product group. For example, if a product in the consumer X product group is not in the consumer B product group, or vice versa, if a product in the consumer B product group is not in the consumer X product group, the product is are not covered by For this reason, the reliability of the correlation coefficient in the conventional art becomes low when the number of products whose purchasing behavior is clear to each other among the consumers who are the targets of the correlation coefficient is small. If the reliability of the correlation coefficient is low, the accuracy of the degree of product recommendation to the consumer X calculated by collaborative filtering naturally decreases.
  • the item recommendation apparatus includes an AI agent that imitates the preferences of the consumer A, an AI agent that imitates the preferences of the consumer B, an AI agent that imitates the preferences of the consumer C, an AI agent that imitates the preferences of the consumer C, a consumer Create an AI agent, . Then, when calculating the correlation coefficient between the consumer X and the consumer A, it is possible to completely match the products of the consumer X product group and the consumer A product group.
  • consumer A's preferences corresponding to consumer X's product group are all the preferences of consumer A, regardless of whether consumer X's product group and consumer A's product group match. is estimated by an AI agent that mimics
  • the AI agent that imitates consumer A's tastes "probably will buy” It can be assumed that there will be no The same can be said for new products that have not been distributed in the market at all. That is, the product recommendation device according to the embodiment of the present invention allows an AI agent that imitates consumer A's preferences to read the image data of a new product that has not yet been distributed in the market. , it can be estimated that consumer A "probably will buy” or "probably will not buy” the product.
  • the item recommendation device can calculate the degree of recommendation of even a new product that has not been distributed in the market at all. Thus, it is possible to solve the cold start problem in prior art collaborative filtering.
  • the article recommendation device clusters consumers to which attribute information such as sex, age, general residence, job title, etc. is linked in advance using the attribute information. Then, based on the consumer's purchase history information, the design image data of the product before it is put on the market is read by an AI agent that imitates the consumer's taste, and an estimated value of purchase possibility is output. As a result, the degree of recommendation for each consumer is calculated by collaborative filtering, and it becomes possible to infer by analogy which group of clustered consumers and how much the product is sold.
  • FIG. 13 is a graph showing estimated results with respect to the data loss rate when confirming the estimated results of the item recommendation device according to the embodiment of the present invention by simulation.
  • the horizontal axis of the graph indicates the missing data rate, and the vertical axis indicates the estimated performance.
  • the broken line shows the result of estimation using actual data.
  • the horizontal straight line indicates the result when estimated using the output of artificial intelligence. When artificial intelligence is used, it is possible to synthesize data, so it is not affected by the loss rate. Therefore, the output of artificial intelligence will be a horizontal straight line. Without artificial intelligence, the lower the defect rate, the better the estimated performance. And regardless of the use of artificial intelligence, the more samples answered, the better the estimated performance.
  • Non-Patent Document 1 describes that an algorithm for specifying an item to be recommended to a customer, which is described as a "recommendation candidate prediction method", can be classified into two types: a memory-based method and a model-based method. Furthermore, it is also shown that there are algorithms that have features of both memory-based and model-based methods.
  • the correlation coefficients described in the previous embodiments belong to memory-based methods. Collaborative filtering based on correlation coefficients maintains a database in which information such as purchase histories and preferences of subjects are recorded in advance. Then, when it comes to the stage of calculating recommendation candidates, a correlation coefficient is calculated by combining the purchase history of the subject and the purchase history of the target customer, which are included in the database. Then, the item recommendation device of the present embodiment calculates the degree of recommendation based on this correlation coefficient, and determines recommendation candidates based on the magnitude of the degree of recommendation. Algorithms such as correlation coefficients are named memory-based methods because they are based on databases. The final stage of this calculation is expressed as "prediction" in Non-Patent Document 1. In addition, hereinafter, the term "taste" of a customer and/or a subject is assumed to be a generic term that also includes purchasing tendency.
  • model-based methods pre-build a "model" before the device that computes the collaborative filtering is used.
  • This model expresses the regularity of product preferences between the customer and the subject, such as "what the customer X likes, the subject A also likes.”
  • the model-based method forms an estimator for estimating the goods that the customer X is likely to buy, based on the preferences such as the purchase histories of the customer X and a plurality of subjects, by performing predetermined arithmetic processing. That is, model-based collaborative filtering is equivalent to forming an approximation function or estimator by a kind of learning algorithm. Algorithms that form models in collaborative filtering are named model-based methods because they are model-based.
  • Non-Patent Document 1 lists cluster models, function models, stochastic models, and time-series models as groups of algorithms that can be used in the model-based method.
  • a cluster model is a modeling of the degree of customer recommendation using clustering, which is a kind of well-known multivariate analysis. That is, the cluster model performs grouping (clustering) on the subject database based on similar tastes, and finds clusters that are closest to the customer's tastes.
  • a function model is obtained based on an approximate function algorithm such as a regression problem, a class classification problem, or an order regression problem, using a utility function or the like that outputs a larger value as a function model matches the customer's preference.
  • the probability model indicates what the approximation function algorithm employed in the function model can interpret as a probability distribution.
  • the probabilistic model is a subset of the functional model.
  • a time-series model is a method that uses changes over time to predict customer preferences, and is a model that adds the concept of changes in customer preferences over time to a cluster model, function model, and/or stochastic model. be.
  • Non-Patent Document 1 discloses various methods as algorithms that can be used in the model-based method. and In other words, it was determined that the regression model is suitable for the algorithms that can be used in the model-based method because of the ease of implementation and the high accuracy of the results obtained. Furthermore, although not disclosed in Non-Patent Document 1, independent component analysis disclosed in Non-Patent Document 3 and Non-Patent Document 4 etc. is an algorithm that can be used in the model-based method adopted for collaborative filtering in the present invention. However, it was judged to be extremely suitable for the high accuracy of the results obtained.
  • FIG. 14 shows the estimated performance of the item recommendation device according to the embodiment of the present invention when confirmed with the Moußs 100K dataset (https://grouplens.org/datasets/moologies/). It is a thing.
  • FIG. 14 shows Youden's J statistics (sensitivity + specificity - 1) obtained by normal collaborative filtering without an AI agent and estimation by collaborative filtering with an AI agent.
  • sensitivity is the probability that you can guess that the estimated grade is positive
  • specificity is the probability that you can guess that the estimated grade is negative. means. In other words, if the estimation results are poor and there is no hit at all, the sum of sensitivity and specificity is "1", and the value of Youden's J statistic (sensitivity + specificity - 1) is infinitely close to "0" be a value.
  • the Moußs100K dataset contains information on 100,000 movie reviews. Here, the data of 200 people who gave the most reviews were picked up, and it was estimated which movies each of them gave a review. First, we created 200 AI agents that take movie images as input and predict whether or not they will give a review (whether they will watch the movie). Then, by collaborative filtering using these, we estimated whether or not each user would leave a review. Since unpopular movies have a low license fee, Netflix (registered trademark) and other video distribution services allow users to watch unpopular movies, thereby improving profitability.
  • FIG. 15 is a block diagram showing software functions of a server 1501 according to a modification of the invention. Differences between the functional blocks of the server 1501 shown in FIG. 15 and the functional blocks of the server 101 shown in FIG. 5 are as follows. (1) A functional model arithmetic processing unit 1503 is provided inside the collaborative filtering arithmetic processing unit 1502 instead of the correlation coefficient arithmetic processing unit 504 . (2) Of the tables read and written by the input/output control unit 501, a target user function model table 1504 is provided instead of the target user correlation coefficient table 510. FIG.
  • FIG. 16 is a diagram showing field configurations of various tables shown in the block diagram of FIG. As described with reference to FIG. 15, among the tables read and written by the input/output control unit 501, instead of the target user correlation coefficient table 510 shown in FIG. 5, the target user function model table 1504 is shown in FIG. .
  • the target user function model table 1504 has a customer ID field, a subject ID group field, and a model parameter field.
  • the customer ID field is the same as the same name field of the target user answer result table 509 .
  • the subject ID group field multiple subject IDs of subjects using the preference data are stored.
  • the order of subject IDs in the subject ID group field is important in calculating model parameters.
  • the subject ID of the subject using the preference data is used for constructing model parameters, which will be described later.
  • the model parameter field stores model parameters for constructing a function model for estimating the customer's preference specified in the customer ID field using the preference data of a plurality of subjects specified in the subject ID group field. be.
  • FIG. 17 is a flow chart showing the work flow (procedure) in the server 1501.
  • FIG. Step S1701 is the same as step S701 in FIG.
  • steps S1702 to S1705 are the same as steps S702 to S705 in FIG.
  • the input/output control unit 501 estimates the favorability ratings for several to several tens of products answered by the customer using the AI agents of the multiple subjects created in step S1703, and stores them in the AI user estimation result table 508 ( S1705). As a result, it is possible to estimate the favorability ratings of a plurality of subjects with respect to several to several tens of products for which customers have answered favorability ratings, making it possible to calculate an approximate function model.
  • the input/output control unit 501 compares the target user answer result table 509 and the AI user estimation result table 508 . Then, the input/output control unit 501 uses the subject's preference data group obtained from the AI user estimation result table 508 as input data and the customer's preference data group obtained from the target user answer result table 509 as labels, and creates a customer function model. Calculate and generate customer model parameters. In most cases, this model parameter is matrix data. Then, the input/output control unit 501 stores this model parameter in the target user function model table 1504 (S1706).
  • the input/output control unit 501 estimates the favorability of the candidate product to be recommended to the customer by the AI agent of the subject used in the calculation of the function model in step S1706. is stored in the AI user estimation result table 508 (S1707).
  • the input/output control unit 501 through the collaborative filtering arithmetic processing unit 503, calculates the degree of recommendation of the product to be recommended to the customer based on the model parameters calculated in step S1706 and the test subject's favorability estimated in step S1707. and stores it in the product recommendation level field of the target user collaborative filtering table 511 (S1708), and the series of processing ends (S1709).
  • FIG. 18A is a schematic diagram showing an outline of the operation of forming a function model that imitates the customer's taste in the function model calculation processing unit 1503.
  • the input/output control unit 501 designates the function model formation mode to the function model calculation processing unit 1503 of the collaborative filtering calculation processing unit 1502, and sets the preference data group of the subject A, the preference data group of the subject B, . is read, and the customer X's preference data group is entered in the label (answer). Then, the function model calculation processing unit 1503 generates a model parameter P1801 that imitates the customer X's preference.
  • FIG. 18B is a schematic diagram showing an overview of the operation of the function model calculation processing unit 1503 when estimating the customer's preference for the target product.
  • the operating state of the function model calculation processing unit 1503 shown in FIG. 18B will be referred to as function model estimation mode.
  • P be an unknown product.
  • the collaborative filtering arithmetic processing unit 1502 designates the function model estimation mode to the function model arithmetic processing unit 1503 and incorporates the customer X's model parameter P1801.
  • the input data includes estimated preference data for subject A's product P estimated by subject A's AI agent, estimated preference data for subject B's product P estimated by subject B's AI agent, and so on.
  • the estimated preference data for the product P of the subject n estimated by the AI agent of the subject n is read. Then, the functional model calculation processing unit 1503 becomes an AI agent that imitates the taste of the customer X, and estimates taste data that imitates the taste of the customer X when he/she sees the image data of the product P.
  • the model for estimating customer X's preferences in collaborative filtering is very similar to the estimator in learning algorithms.
  • Collaborative filtering performed by the server 101 and the server 1501 of the present invention inserts into the input data for estimating the preferences of the customer X, presumed data of preferences for a given product by an AI agent that imitates the preferences of the subject.
  • it is a big difference from the conventional technology and also an advantage.
  • Non-Patent Document 1 collaborative filtering based on correlation coefficients was explained as a separate category from models such as functions. , there is not much difference between the correlation coefficient (memory-based method) and the function model (model-based method). In the case of the correlation coefficient, the function model formation mode of FIG. 18A and the function model estimation mode of FIG. 18B are almost unified. Corresponding to the model parameters are the correlation coefficient and the coefficient for the preference data based on the correlation coefficient.
  • the function model calculation processing unit 1503 implemented in the embodiment of the present invention generates a customer preference model, which is a function model for estimating the customer's preference from the estimated value of the subject's preference, according to a predetermined calculation procedure. It can also be said that it is a customer preference model formation processing unit.
  • the customer preference model formation processing unit also includes arithmetic processing based on the aforementioned correlation coefficients.
  • the item recommendation device includes a neural network operation processing unit (NN operation) that imitates the preferences of a plurality of subjects and outputs an estimated value of the subject's preference for input image data of an item.
  • a processing unit 502) and a customer preference model formation processing unit (correlation coefficient calculation processing unit 504 and function model arithmetic processing unit 1503) and the customer preference model read the image data of a plurality of items, estimate the customer's preference for the plurality of items, calculate the degree of recommendation of the item to the customer, and calculate the degree of recommendation.
  • a collaborative filtering arithmetic processing unit collaborative filtering arithmetic processing unit 503 and collaborative filtering arithmetic processing unit 1502 that selects articles to be recommended to the customer based on the above.
  • an item recommendation device has been described.
  • the article recommendation device of this embodiment makes AI learn purchasing behavior data or preference data of a plurality of subjects.
  • an AI agent that has learned the preferences of each subject and imitates the preferences of each subject is created.
  • the AI agent estimates the subject's preference for unknown products when calculating the correlation coefficient with the customer.
  • the product recommendation device of the present embodiment recommends the optimum product to the customer based on the purchase behavior data or preference data of subjects with a high correlation coefficient with the customer, that is, subjects with similar purchase behavior or preferences.
  • the AI agent that imitates each subject's preference estimates the subject's preference for the item related to the customer's purchase history. can be done. Therefore, it is possible to maximize the number of elements used for calculation of the correlation coefficient based on the customer's purchase history. Furthermore, even for an unknown product that has not been distributed in the market at all, it is possible to estimate the subject's preference as long as there is image data, so it is possible to predict how much the product will sell.
  • Operation unit 208 Bus 301 CPU 302 ROM 303 RAM 304 Nonvolatile storage 305 NIC 306 Display unit 307 Operation unit 308 Bus 401 Input/output control unit 501 Input/output Control unit 502 NN operation processing unit 503 Collaborative filtering operation processing unit 504 Correlation coefficient operation processing unit 505 Product master 506 AI user coefficient table 507 AI user answer result table 508 AI User estimation result table 509 Target user answer result table 510 Target user correlation coefficient table 511 Target user collaborative filtering table 1501 Server 1502 Collaborative filtering operation processing unit 1503 Function model operation processing unit 1504 ... Target user function model table

Landscapes

  • Business, Economics & Management (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Engineering & Computer Science (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne un dispositif permettant de sélectionner des produits qui correspondent aux préférences d'un client et/ou à une tendance à l'achat avec une précision d'inférence plus élevée par comparaison avec une précision classique, et de recommander les produits au client. Lorsque le dispositif amène un agent IA à apprendre des données de comportement d'achat ou des données de préférence d'une pluralité de sujets de test, l'agent IA qui a appris les préférences des sujets de test respectifs, et imite les préférences des sujets de test respectifs, est créé. Puis, l'agent IA infère les préférences des sujets de test par rapport à un produit inconnu lors du calcul d'un coefficient de corrélation avec un client, et recommande des produits optimaux au client à partir des données de comportement d'achat ou des données de préférence de sujets de test ayant un coefficient de corrélation élevé avec le client, c'est-à-dire des sujets de test ayant un comportement ou des préférences d'achat très similaires.
PCT/JP2022/040787 2021-11-12 2022-10-31 Dispositif de recommandation d'article et procédé de recommandation d'article WO2023085165A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2023559578A JPWO2023085165A1 (fr) 2021-11-12 2022-10-31

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2021184596 2021-11-12
JP2021-184596 2021-11-12

Publications (1)

Publication Number Publication Date
WO2023085165A1 true WO2023085165A1 (fr) 2023-05-19

Family

ID=86335954

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2022/040787 WO2023085165A1 (fr) 2021-11-12 2022-10-31 Dispositif de recommandation d'article et procédé de recommandation d'article

Country Status (2)

Country Link
JP (1) JPWO2023085165A1 (fr)
WO (1) WO2023085165A1 (fr)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002334257A (ja) * 2001-05-10 2002-11-22 Nec Corp レコメンドエンジン、レコメンド方法、レコメンドプログラム
WO2016016934A1 (fr) * 2014-07-29 2016-02-04 株式会社日立製作所 Système d'analyse des préférences
JP2016048417A (ja) * 2014-08-27 2016-04-07 石井 美恵子 サービス提供システムおよびプログラム
US11037222B1 (en) * 2018-08-27 2021-06-15 A9.Com, Inc. Dynamic recommendations personalized by historical data

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002334257A (ja) * 2001-05-10 2002-11-22 Nec Corp レコメンドエンジン、レコメンド方法、レコメンドプログラム
WO2016016934A1 (fr) * 2014-07-29 2016-02-04 株式会社日立製作所 Système d'analyse des préférences
JP2016048417A (ja) * 2014-08-27 2016-04-07 石井 美恵子 サービス提供システムおよびプログラム
US11037222B1 (en) * 2018-08-27 2021-06-15 A9.Com, Inc. Dynamic recommendations personalized by historical data

Also Published As

Publication number Publication date
JPWO2023085165A1 (fr) 2023-05-19

Similar Documents

Publication Publication Date Title
Han et al. A fuzzy TOPSIS method for performance evaluation of reverse logistics in social commerce platforms
Cheng et al. A fuzzy recommender system based on the integration of subjective preferences and objective information
Scholz et al. Using PageRank for non-personalized default rankings in dynamic markets
KR102321987B1 (ko) 온라인 쇼핑몰 관리 시스템 및 방법
CN114219169A (zh) 颖幡供应链销售和库存预测算法模型和应用系统
KR20160065429A (ko) 하이브리드 개인화 상품추천 방법
Yan et al. Implementation of a product-recommender system in an IoT-based smart shopping using fuzzy logic and apriori algorithm
Zaim et al. Multi-criteria analysis approach based on consumer satisfaction to rank B2C E-commerce websites
Chan et al. IoT data acquisition in fashion retail application: Fuzzy logic approach
Ahamed et al. A recommender system based on deep neural network and matrix factorization for collaborative filtering
Bhattacharjee et al. An integrated machine learning and DEMATEL approach for feature preference and purchase intention modelling
CN117495482A (zh) 一种基于用户画像的二手手机销售推荐方法及系统
WO2023085165A1 (fr) Dispositif de recommandation d'article et procédé de recommandation d'article
Akhtarzada et al. A multi-criteria metric algorithm for recommender systems
Sharma et al. Recommendation system for movies using improved version of som with hybrid filtering methods
Kim et al. A New Perspective for Neural Networks: Application to a Marketing Management Problem.
Nune Characterizing and predicting early reviewers for effective product marketing on ecommerce websites
Stephen et al. An intelligent cloud-based customer relationship management system to determine flexible pricing for customer retention
Ivanov et al. Big Data based marketing forecasting.
Onokoy et al. Modern approaches to building recommender systems for online stores
Ji et al. Personalized recommendation based on a multilevel customer model
Liu et al. Improving Recommendation Accuracy by Considering Electronic Word-of-Mouth and the Effects of Its Propagation Using Collective Matrix Factorization
Maurya et al. Analysis of behavioural data of customer for the E-commerce platform by using machine learning approach
Umeda et al. Evaluation of collaborative filtering by agent-based simulation considering market environment
Huang et al. Research on consumer psychology modeling and intelligent recommendation of e-commerce users

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22892664

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2023559578

Country of ref document: JP

NENP Non-entry into the national phase

Ref country code: DE