Information processing apparatus and information processing method for predicting preference of user group

Abstract

To predict a preference of a group to which users belong more appropriately with higher accuracy based on preferences of the users for items.SOLUTION: In a group recommendation system, a group recommendation device 1 comprises: an encoding unit for encoding a behavior history of a user for an item to generate a first embedding representation for each user; a weighting unit for deriving a weight for each user based on information on a price of the item and the behavior history, and weighting the first embedding representation generated by the encoding unit with the derived weight; and an aggregation unit for aggregating the first embedding representation weighted by the weighting unit to generate, for a group to which the user belongs, a second embedding representation indicating a preference of the group for the item.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing device and an information processing method for predicting the preferences of a user group, and in particular to a technology for recommending items to multiple users belonging to the group in accordance with the predicted preferences of the user group. [Background technology]

[0002] In the field of electronic commerce (EC) platforms, many EC sites have been built on the web, and many items such as products and services are routinely made available for users to browse. Many of these EC sites have implemented a recommendation system that displays recommended items that a user is likely to purchase based on the user's past purchase history and attribute information, alongside the screen the user is currently viewing. Personalized item recommendations attract more user interest and contribute significantly to the user's decision to purchase the item, thereby increasing advertising effectiveness.

[0003] Group recommendation is one method of recommending such items. In group recommendation, a large number of users who use an e-commerce site are segmented, and items that are more suitable for each segmented user group are individually recommended to that user group. By providing different optimized advertising campaigns to different user groups, it is expected that the effectiveness of digital marketing for all users belonging to the user group will be improved.

[0004] In this group recommendation, how accurately the preferences of a user group consisting of a large number of users can be predicted will have a major impact on sales of the item in question and related items on the e-commerce site.

[0005] Non-Patent Document 1 discloses a group recommendation system that predicts the preferences of an entire user group for an item by aggregating preferences for the item of multiple users belonging to the user group.

[0006] Specifically, the group recommendation system disclosed in Non-Patent Document 1 uses a neural network to derive the preferences of multiple users belonging to a user group, aggregates the derived preferences of the multiple users, derives preferences that represent the user group, and predicts items that best match the derived preferences of the user group. The group recommendation system disclosed in Non-Patent Document 1 further uses a classifier network that has been pre-trained to more closely distinguish between preferences for items between users who are members of a user group and users who are not members of the user group. Using this classifier network, the group recommendation system disclosed in Non-Patent Document 1 assigns weights to the preferences of each user who belongs to the user group, aggregates the weighted preferences of each user, and derives preferences that represent the user group. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Aravind Sankar, et. al., “GroupIM: A Mutual Information Maximization Framework for Neural Group Recommendation”, Proceedings of the 43rd International Association for Computing Machinery's Special Interest Group on Information Retrieval (ACM SIGIR) Conference on Research and Development in Information Retrieval, pp.1279-1288,July,2020. Summary of the Invention [Problem to be solved by the invention]

[0008] The technology in Non-Patent Document 1 aims to identify users who have more influence on decision-making in the user group in comparison with other users when deriving the preferences of the entire user group, and assign a higher weight to the identified users. As a result, the preferences of users who are assigned a higher weight are more likely to be reflected in the preferences of the user group.

[0009] However, the number of users who use e-commerce sites on e-commerce platforms is extremely large, and their attributes are diverse. Here, it is relatively easy to predict the preferences of a user group formed by permanent relationships such as family, friends, or colleagues, because the users are closely related to each other and often take the same actions.Even in a user group formed by temporary relationships such as participants in a conference or meeting, it is easy to predict common preferences because they at least temporarily share time and place and are interested in common interests.

[0010] In contrast, the large number of users who use e-commerce sites are usually segmented based on a small number of user attributes, such as the user's age group or residential area, and many users are geographically dispersed. Therefore, unlike the above-mentioned user groups in the real world, user groups made up of e-commerce site users inherently have little in common with each other in terms of preferences.

[0011] Furthermore, unlike the real-world user groups, such user groups consisting of e-commerce site users have weak direct correlations between users, and there is almost no history of users performing the same actions at the same time. This makes it more difficult to identify users with greater influence within a user group, which may reduce the accuracy of predicting the preferences of a user group. In particular, in large-scale e-commerce platforms, when a user group is made up of a large number of users who lack interrelationships, there is a high possibility that items recommended to that user group according to the predicted preferences of that user group will lack the appeal to stimulate users' desire to purchase.

[0012] The present invention has been made to solve the above-mentioned problems, and its purpose is to provide an information processing device and an information processing method that can more appropriately and accurately predict the preferences of a group to which a user belongs based on the user's preferences for items. [Means for solving the problem]

[0013] In order to solve the above problem, one aspect of the information processing device of the present invention includes an encoding unit that encodes a user's behavioral history regarding an item and generates a first embedded representation for each user; a weighting unit that derives a weight for each user based on information regarding the price of the item and the behavioral history and weights the first embedded representation generated by the encoding unit with the derived weight; and an aggregation unit that aggregates the first embedded representations weighted by the weighting unit and generates, for a group to which the user belongs, a second embedded representation that indicates the group's preference for the item.

[0014] One aspect of the information processing method of the present invention is an information processing method executed by an information processing device, comprising the steps of: encoding a user's behavioral history regarding an item to generate a first embedded representation for each user; deriving a weight for each user based on information regarding the price of the item and the behavioral history, and weighting the generated first embedded representation with the derived weight; and aggregating the weighted first embedded representations to generate a second embedded representation for a group to which the user belongs, the second embedded representation indicating the group's preference for the item.

[0015] One aspect of the computer-readable recording medium according to the present invention is a computer-readable recording medium storing an information processing program for causing a computer to execute information processing, the program causing the computer to perform the following steps: an encoding process for encoding a behavior history of a user with respect to an item to generate a first embedded representation for each user; and a weighting process for deriving a weight for each user based on information related to the price of the item and the behavior history, and weighting the first embedded representation generated by the encoding process with the derived weight. and an aggregation process of aggregating the first embedded representations weighted by the weighting process to generate, for a group to which the user belongs, a second embedded representation that indicates the group's preference for the item. [Effects of the Invention]

[0016] According to the present invention, it is possible to more appropriately and accurately derive the preferences of groups to which a user belongs from the preferences of the user for items. The above-mentioned objects, aspects, and advantages of the present invention, as well as other objects, aspects, and advantages of the present invention not described above, will be understood by those skilled in the art from the following detailed description of the invention by referring to the accompanying drawings and the claims. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a block diagram showing an example of the functional configuration of a group recommendation system according to each embodiment of the present invention. [Figure 2] FIG. 2 is a conceptual diagram showing an example of group recommendation in which the function of deriving preferences of a user group in the previous stage is implemented in a first neural network equipped with an attention mechanism. [Figure 3] FIG. 3 is a conceptual diagram showing an example of group recommendation in which a function for recommending items to a subsequent user group is implemented in the second neural network. [Figure 4] FIG. 4 is a flowchart showing an example of an outline of the processing procedure of the group recommendation processing executed by the group recommendation device 1 according to this embodiment. [Figure 5] FIG. 5 is a flowchart showing an example of detailed processing procedures of the weighting process in step S3 of FIG. 4, which is executed by the group recommendation device 1 according to this embodiment. [Figure 6] Figure 6 is a diagram showing an example of the relationship between the input item price, the user's item purchase frequency, and the output weight in the adaptive sigmoid function implemented in the user identification machine learning model 31 in this embodiment. [Figure 7]Figure 7 is a conceptual diagram showing an example of the processing procedure of a process executed by the group recommendation device 1 according to this embodiment to derive preferences of a user group using a machine learning model 31 for user identification and a machine learning model 32 for generating group embedded expressions. [Figure 8] FIG. 8 is a conceptual diagram showing an example of a processing procedure of a process for recommending an item to a user group using the item recommendation machine learning model 33, which is executed by the group recommendation device 1 according to this embodiment. [Figure 9] FIG. 9 is a flowchart showing an example of detailed processing procedures for pre-learning each machine learning model according to this embodiment. [Figure 10] FIG. 10 is a block diagram showing an example of the hardware configuration of the group recommendation device 1 according to this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, with reference to the accompanying drawings, an embodiment for carrying out the present invention will be described in detail. Among the components disclosed below, those having the same function will be given the same reference numerals, and their description will be omitted. Note that the embodiment disclosed below is an example of a means for realizing the present invention, and should be appropriately modified or changed depending on the configuration of the device to which the present invention is applied and various conditions. The present invention is not limited to the following embodiment. Furthermore, not all of the combinations of features described in the present embodiment are necessarily essential to the solution of the present invention.

[0019] The group recommendation device of this embodiment encodes a user's behavioral history regarding items, extracts the user's preferences regarding items for each user, derives a weight indicating the importance of the user within the user group using the item information for each user, weights the extracted user preferences with the derived weight, and aggregates the weighted preferences to generate preferences for the user group to which the user belongs. The group recommendation device according to this embodiment may further predict a score indicating the preferences of the user group for items from the generated user group preferences and information on the corresponding items. The group recommendation device according to this embodiment may further predict and output items to be recommended to the user group from the predicted preferences of the user group for items.

[0020] Below, we will explain a non-limiting example in which the group recommendation device of this embodiment generates embedded expressions (embeddings) representing a user's preferences for items from the user's purchase history and other operation history on e-commerce sites and other sites that allow users to purchase products and services as the user's behavioral history, and aggregates multiple embedded expressions for multiple users belonging to the same user group to generate embedded expressions representing the group's preferences for items.

[0021] However, this embodiment is not limited to this. For example, user preferences may be derived based on other behavioral history, such as transaction history at a real store or facility, rather than an e-commerce site. Furthermore, information representing a user's preferences for an item is not limited to embedded representations, and may be in any format as long as it can be used as a feature in machine learning processing. Here, "preferences" broadly refers to information that associates a user with an item, which can be understood through the user's behavior related to the item. Furthermore, "behavioral history" here includes any behavioral history related to an item, such as item purchase history, favorite item setting history, and item click history. Furthermore, a user group may be a cluster obtained by segmenting multiple users based on any attribute of the user that may be related to the purchase of an item or other transactions, such as the user's age group, residential area, annual income group, family structure, or analyzed persona. There is no limit to the number of users that belong to a certain user group, but a user group is typically made up of a large number of users.

[0022] In the following, an example will be described in which the group recommendation device according to this embodiment aggregates preferences for items of multiple users to generate preferences for a user group, and infers and outputs items to be recommended to the user group based on the generated preferences of the user group. However, this embodiment is not limited to this, and may include an aspect in which inference and output of items to be recommended to the user group are not performed. In this embodiment, the term "item" broadly includes goods and services that can be traded by users.

[0023] <Functional configuration of group recommendation system> 1 is a block diagram showing an example of the functional configuration of a group recommendation system according to this embodiment. The group recommendation system shown in FIG. 1 includes a group recommendation device 1, a transaction information storage device 2, and a machine learning model storage device 3. The group recommendation device 1 includes an input unit 11, an encoding unit 12, a weighting unit 13, an aggregation unit 14, a group preference prediction unit 15, and a recommended item output unit 16. The group recommendation device 1 may further include a pre-learning unit 17, but instead, the pre-learning unit 17 may be provided in a computing component other than the group recommendation device 1.

[0024] The transaction information storage device 2 is a storage device that stores information related to transactions between the EC site and the user, including EC site usage history 21 and item information 22. The EC site usage history 21 includes a history of items purchased or viewed by the user using the EC site in the past, and attribute information of the user who used the EC site. The item information 22 includes attribute information of the item.

[0025] The machine learning model storage device 3 is a storage device that stores a machine learning model for user identification 31, a machine learning model for generating group embedded expressions 32, and a machine learning model for recommending items 33. The machine learning model for user identification 31, the machine learning model for generating group embedded expressions 32, and the machine learning model for recommending items 33 may be machine learning models configured separately, or may be configured as one or two machine learning models by incorporating one into the other as a module or by pipeline coupling. The transaction information storage device 2 and the machine learning model storage device 3 may each be configured as a non-volatile storage device such as a hard disk drive (HDD) or a solid state drive (SSD). Furthermore, the transaction information storage device 2 and the machine learning model storage device 3 only need to be configured to be accessible from the group recommendation device 1, and may be provided in other computing components.

[0026] The group recommendation device 1 may also be communicably connected to a client device (not shown) such as a PC (Personal Computer) via a network. In this case, the group recommendation device 1 is implemented in a server, and the client device may provide a user interface for the group recommendation device 1 to input and output information from and to the outside, and may also include some or all of the components 11 to 17 of the group recommendation device 1.

[0027] The input unit 11 acquires the EC site usage history 21 and item information 22 stored in the transaction information storage device 2 as input data, and supplies the acquired EC site usage history 21 and item information 22 to the encoding unit 12 and the weighting unit 13. Here, the EC site usage history 21 may include, as information representing a user's preferences for items, all behavioral history including a history of operations performed by the user related to items on the EC site, attributes that characterize each user, information on the user group to which the user belongs, etc. The item information 22 may include information indicating whether each user has purchased the item, various attributes that characterize the item, such as the price, category, and brand of the item, and an image of the item.

[0028] The input unit 11 extracts, from the EC site usage history 21, behavioral histories for target items of one or more users belonging to the same user group who have behavioral histories for target items for which the preferences of the user group are to be predicted for group recommendation processing, and supplies the extracted behavioral history information to the encoding unit 12 and the weighting unit 13. The input unit 11 also supplies, from the item information 22, item information for target items for which the preferences of the user group are to be predicted for group recommendation processing to the encoding unit 12 and the weighting unit 13. The input unit 11 may acquire input data to be processed by reading out the EC site usage history 21 and item information 22 stored in advance in the transaction information storage device 2, or may receive the EC site usage history 21 and item information 22 via a communication I / F from the same or a different counterpart device that stores the EC site usage history 21 and item information 22. Alternatively, the input unit 11 may acquire and update the EC site usage history 21 and item information 22 in real time using a technique such as web crawling via the communication I / F. The input unit 11 also accepts input of various parameters necessary for executing the group recommendation process in the group recommendation device 1. The input unit 11 may accept input of various parameters via a user interface of a client device communicatively connected to the group recommendation device 1.

[0029] The encoding unit 12 encodes the user behavior history information and item information 22 supplied from the input unit 11, and supplies the encoded user behavior history information and item information 22 to the weighting unit 13 and the aggregation unit 14, respectively. Specifically, the encoding unit 12 converts the user behavior history information and the item information 22 into an embedded representation of the user behavior history information and an embedded representation of the item information 22, respectively. The encoding unit 12 generates the embedded representation by converting each component of the input data into a multidimensional feature vector representation that is mapped to a multidimensional vector space.

[0030] The encoding unit 12 generates, for each user or users belonging to a user group, an embedded representation of user behavior history information, and also generates an embedded representation of item information 22 of target items that have interacted with the user group, i.e., for which any user in the user group has a behavior history such as a purchase history. The embedded representation of the user behavior history information is an embedded representation that indicates each user's preference for the target item. Note that, in the following, an example is described in which the group recommendation device 1 generates multiple embedded representations for each of multiple users. However, if there is only one user in the user group who has a behavior history for the target item, a single embedded representation may be generated for that user, and the same applies to the processing of the weighting unit 13 and the aggregating unit 14 described below. In this embodiment, the EC site usage history 21 input to the input unit 11 includes at least the target user's behavioral history, including the purchase frequency of the target user's past purchases of any of a plurality of items offered on the target EC site. This is hereinafter referred to as the "item purchase frequency." Therefore, the embedded representation of the user behavioral history information generated by the encoding unit 12 includes a feature of the purchase frequency of the target user's past purchases of items on the EC site.

[0031] The item information 22 input to the input unit 11 includes at least information on the price of each item, and therefore the embedded representation of the item information 22 generated by the encoding unit 12 includes a feature of the price of the target item. Hereinafter, the embedded representation of the user behavior history information will be referred to as the "user embedded representation," and the embedded representation of the item information will be referred to as the "item embedded representation." This embedded representation encodes the unique properties of the entity being processed. In other words, the user embedded representation encodes at least the user's past behavior history with respect to the target item as interactions with the target item. Furthermore, the item embedded representation encodes various attributes related to transactions of the item on the e-commerce site.

[0032] The weighting unit 13 calculates a weight for each of the user-embedded expressions of multiple users supplied from the encoding unit 12, assigns the calculated weight to each user-embedded expression, and supplies the weighted user-embedded expressions to the aggregation unit 14. Specifically, the weighting unit 13 uses a machine learning model 31 for user identification to input the user embedded representation and the corresponding item embedded representation supplied from the encoding unit 12 into the machine learning model 31 for user identification, and assigns the weight for each user output from the machine learning model 3 for user identification to the corresponding user embedded representation. In this embodiment, the weight of a user calculated by the weighting unit 13 indicates the degree of influence the user has on item purchases within the user group. That is, a user with a larger weight assigned to him / her will contribute more to generating the embedded representation of the user group output by the aggregation unit 14 in the subsequent stage, and will therefore have a more dominant influence on the formation of preferences regarding item purchases for the user group.

[0033] The user identification machine learning model 31 is a classifier constructed by a neural network or algorithm equipped with an attention mechanism for identifying each user belonging to a user group. The attention mechanism includes an encoder and a decoder, and calculates a score from the inner product of the KEY vector of the encoder and the QUERY vector of the decoder. In this embodiment, the user identification machine learning model 31 uses this attention mechanism to generate an attention weight for each user from the user embedding representation and the item embedding representation. This attention weight indicates the influence or importance of each of multiple users in predicting the preferences of the user group and, ultimately, in predicting recommended items. The weighting unit 13 may generate the attention weight using the user identification machine learning model 31, or the weighting unit 13 itself may be configured by the user identification machine learning model 31.

[0034] In this embodiment, the weighting unit 13 calculates a weight to be assigned to the user-embedded expression for each of multiple users belonging to the same user group based on the correlation between the user's item purchase frequency and the item price from the user-embedded expression including the item purchase frequency and the item-embedded expression including the item price, the details of which will be described later with reference to Figures 5 and 6. In this embodiment, the user identification machine learning model 31 may be pre-trained using a training dataset that has been prepared in advance.

[0035] The aggregation unit 14 aggregates the weighted user embedded expressions of multiple users belonging to the same user group supplied from the weighting unit 13, and generates an embedded expression of preferences for a target item that represents the user group. Hereinafter, the aggregated embedded expression of the user group will be referred to as a "group embedded expression." Specifically, the aggregating unit 14 calculates a weighted sum of multiple user-embedded expressions according to weights assigned to each of multiple users belonging to the same user group, and aggregates the multiple user-embedded expressions into a single group-embedded expression that represents the group. Note that the aggregating unit 14 may also aggregate the multiple user-embedded expressions into a group-embedded expression by applying a predetermined aggregation function to the multiple user-embedded expressions instead of the weighted sum.

[0036] A typical method for deriving the preferences of a user group involves directly deriving preferences from the user group's past interactions with a target item and treating the user as a virtual user. However, in the case of a user group on an e-commerce site, there is no uniform behavior as a user group, so there is almost no interaction with the target item. As a result, the data on interactions between the user group and the target item input into the machine learning model is extremely sparse and ineffective. In contrast, in this embodiment, the user group's preferences for the target item are derived by aggregating the past interactions of users belonging to the user group with the target item. In addition, in this embodiment, the aggregation unit 14 aggregates multiple user embedded expressions into a single group embedded expression in the same vector space (embedding space) as the user embedded expressions using a representation learning method that uses attention weights learned by the user identification machine learning model 31 to learn feature expressions of preferences for items that represent user groups.

[0037] The group preference prediction unit 15 predicts the preference of the user group for the item from the group embedded expressions aggregated by the aggregation unit 14, the corresponding item embedded expressions, and the group embedded expressions unique to the group. Specifically, the group preference prediction unit 15 connects the group embeddings with the corresponding item embeddings and propagates the connected group-item embeddings through a multi-layered neural network to predict the score of the target item for the target user group. This score quantitatively indicates the degree of preference of the target user group for the target item and is provided to the subsequent recommended item output unit 16.

[0038] The recommended item output unit 16 outputs items to be recommended to the user group based on the scores for the items of the user group supplied from the group preference prediction unit 15. Specifically, the recommended item output unit 16 may output the item as a recommended item if the score for the item of the user group supplied from the group preference prediction unit 15 is equal to or greater than a predetermined preference threshold that has been set or learned in advance. Alternatively, the recommended item output section 16 may hold a plurality of scores for a plurality of items for the user group, sort the held plurality of items in order of score, and output the list of recommended items. The recommended items may be output to a device such as a display device in a manner that allows the user to view them.

[0039] The pre-learning unit 17 pre-trains the user identification machine learning model 31. When the group recommendation device 1 according to this embodiment is implemented with one or more machine learning models, the pre-learning unit 17 may pre-train each of the machine learning models. These machine learning models have mutually different structures and are pre-trained using different training datasets, the details of which will be described later with reference to FIG. 9. However, in this embodiment, the machine learning model that can be used for the group recommendation process is not limited to a supervised machine learning model, and for example, an unsupervised machine learning model based on reinforcement learning or the like may also be used.

[0040] <Network configuration for group embedding generation and item recommendation processing> Figure 2 is a conceptual diagram showing an example in which the functions of the group embedding expression generation process executed by the upstream input unit 11, encoding unit 12, weighting unit 13, and aggregation unit 14 in the group recommendation device 1 are implemented in a first network. FIG. 3 is a conceptual diagram showing an example in which the functions of the item recommendation process executed by the group preference prediction unit 15 and the recommended item output unit 16 in the subsequent stages in the group recommendation device 1 are implemented in a second network.

[0041] Referring to FIG. 2, the first network comprises a neural network including an input layer 201, an embedding layer 202, a pooling layer 203, an attention mechanism 204, and an output layer 205. The first network is a representation learning model that learns a single group embedding representation to be mapped to the same vector space from multiple user embedding representations arranged in a vector space, and has a neural network structure.

[0042] An index list 201a of users belonging to a certain user group and an index list 201b of items are input to the input layer 201 of the first network. All users belonging to the same user group may be input to the input layer as the user index list 201a. The user behavior history information of the user input to the input layer 201 includes the behavior history of the target user with respect to the target item and various attributes of the user as information representing the target user's past preferences for items with which the target user has interacted.

[0043] Also in FIG. 2, an item index list 201b is entered in which one item from a list of related items is entered as a target item for which preferences of a user group are to be inferred. The item information of an item input to the input layer 201 includes at least information on whether the item has been purchased by any user in the user group, various attributes of the item, and an image of the item. The item attributes include at least the price of the item.

[0044] The list of users 201a and the list of items 201b input to the input layer 201 are encoded in the embedding layer 202 of the first network to generate a user embedding representation u for each user. kj,1 ,u kj,2 ,u kj,3 and the item embedding representation of each item v j and placed in the vector space. The user embedding representation u generated by the embedding layer 202 kj,1 ,u kj,2 ,u kj,3 In the pooling layer 203 of the first network, attention weights are assigned by the attention mechanism 204, which is a neural attention mechanism, and the resulting representations are converted into weighted user embedded representations α(j,1),α(j,2),α(j,3).

[0045] Specifically, the attention mechanism 24 calculates the item embedding representation v j Using this, we can calculate the user embedding representation u kj,1 ,u kj,2 ,u kj,3 The attention mechanism 24 then derives attention weights α(j,1), α(j,2), and α(j,3) for each user. These attention weights identify each user within the user group l and function as indicators of the degree of influence of each user on the user group's decision-making regarding the purchase of the target item j. These attention weights may vary for each user depending on the past preferences indicated by each user's behavioral history and the attributes of the target item. The attention mechanism 24 also applies the derived attention weights α(j,1), α(j,2), and α(j,3) to the user embedding representation u of each user, respectively. kj,1 ,u kj,2 ,u kj,3By adding , it is converted into a weighted user embedding representation for each user.

[0046] These weights α(j,1), α(j,2), α(j,3) are used to generate the user embedding representation u in the machine learning model for group embedding representation generation 32. kj,1 ,u kj,2 ,u kj,3 and the item embedding representation v j is input, and is a parameter α(j,t) that can be learned by the attention mechanism 204. In this embodiment, the machine learning model for user identification 31 performs the function of the attention mechanism 204 in Fig. 2, the details of which will be described later with reference to Fig. 6. The pooling layer 203 further aggregates the weighted user embeddings of multiple users to generate a group embedding 205a representing the user group, and outputs the group embedding 205a to the outside via the output layer 205. Note that the user group itself may or may not have an interaction with item j. If the user group has an interaction with item j, an embedded representation indicating the user group's own preference for the item can be directly generated. Therefore, the generated embedded representation of the user group may be added to the group embedded representation 205a obtained by aggregating the user embedded representations of multiple users. For example, if at least multiple users in a user group have an interaction with the target item, the user group may be considered to have an interaction with the target item.

[0047] Referring to FIG. 3, the second network comprises a neural network including a representation layer 301, a pooling layer 302, a hidden layer 303, and a prediction layer 304. The second network generates a group embedding representation g that indicates the preferences of user group l for target item j, which is output by the first network. l (j) and the embedding representation of item j, v j From this, we learn the interactions between user groups and items, and calculate the score ^y of item j in user group l.lj The learning model may have a multi-layer neural network structure such as NCF (Neural Collaborative Filtering). The first network and the second network may be implemented in a single machine learning model, or may form a pipeline together.

[0048] The representation layer 301 of the network in Figure 3 contains a group embedding representation g that indicates the preference for item j. l (j) and the embedding representation v of the target item j j is entered. In the pooling layer 302 of the second network, the group embedding representation g l (j) and the embedding representation of item j, v j In this pooling layer 302, the product of the calculated product and the group embedding representation g l (j) and the embedding representation v of the target item j j The group-item concatenated embedded representation 302a is generated by concatenating the group-item and the group-item concatenated embedded representation g. l (j) and the embedding representation of target item j, v j It is a vector representation that models the interaction with

[0049] The group-item concatenated embedded representations 302a generated in the pooling layer 302 are propagated to multiple layers 303a that make up the hidden layer 303 in the second network. The prediction layer 304 of the second network uses the scores ^y output from the hidden layer 303. lj is output as a predicted value indicating the degree of purchase willingness or preference for item j of user group l.

[0050] The second network shown in Figure 3 can further predict preferences for item j for each user. Referring to FIG. 3, similar to the case of user groups, the representation layer 301 of the second network contains user embedding representations u that indicate the user's preferences for item j. i and the embedding representation v of the target item j j is entered. In the pooling layer 302 of the second network, the user embedding representation u i and the embedding representation v of the target item j j In this pooling layer 302, the product of the calculated product and the user embedding representation u i and the embedding representation v of the target item j j The user-item concatenated embedded representation 302b is generated by concatenating the user-item concatenated embedded representation u i and the embedding representation v of the target item j j It is a vector representation that models the interaction with

[0051] The user-item connected embedded representation 32b generated in the pooling layer 302 is propagated to multiple layers 303a constituting the hidden layer 303 of the second network. The multiple hidden layers 303 generate a score ^y lj may be shared for calculation purposes. The prediction layer 304 of the second network uses the scores ^r output from the hidden layer 303. ij is output as a predicted value indicating the degree of purchase willingness or preference for the target item j of each user. The second network calculates the score ^r for each user's target item j. ij In this case, the representation layer 301 of the second network may be configured not to calculate the user embedding representation u for each user. i does not need to be entered.

[0052] <Group recommendation process executed by the group recommendation device 1> 4 is a flowchart showing an example of an outline of the processing steps of the group recommendation processing executed by the group recommendation device 1 according to this embodiment. The group recommendation processing shown in FIG. 4 corresponds to an inference phase in which the group recommendation device 1 executes the processing using the user identification machine learning model 31, the group embedded expression generation machine learning model 4, and the item recommendation machine learning model 33 to predict the preferences of a user group. Each step in FIG. 4 is realized by a CPU reading and executing a program stored in a storage device such as an HDD of the group recommendation device 1. At least a part of the flowchart shown in FIG. 4 may be realized by other hardware such as a GPU. When realized by hardware, for example, a predetermined compiler may be used to automatically generate a dedicated circuit on an FPGA (Field Programmable Gate Array) from a program for realizing each step. Alternatively, a gate array circuit may be formed in the same manner as an FPGA and realized as hardware. Alternatively, it may be realized by an ASIC (Application Specific Integrated Circuit).

[0053] In step S1, the input unit 11 of the group recommendation device 1 acquires the EC site usage history 21 and item information 22 for one or more users belonging to the same user group from the transaction information storage device 2, and supplies them to the encoding unit 12 and the weighting unit 13. Here, the EC site usage history 21 is information for deriving a user's past preferences for items, and may include all behavioral history including all operations performed by the user in relation to items on the EC site, attributes characterizing each user, information on the user group to which the user belongs, etc. The item information 22 may include information indicating whether each user has purchased the item, various attributes characterizing the item, an image of the item, etc.

[0054] In this embodiment, information on the purchase frequency of an item on a target EC site and the price of the target item included in the user's behavior history of the EC site usage history 21 is used to derive an attention weight to be assigned to the target user. For this reason, the EC site usage history 21 input to the input unit 11 in step S1 includes at least the user's purchase frequency of the target item and other items on the EC site, and the item information 22 input to the input unit 11 in step S1 includes at least information on the price of the target item.

[0055] In step S2, the encoding unit 12 of the group recommendation device 1 encodes the EC site usage history 21 supplied from the input unit 11 in step S1, maps it to a vector space, and generates a user embedded representation u kj,1 ,u kj,2 ,u kj,3 Similarly, the encoding unit 12 encodes the item information 22 provided by the input unit 11 in step S1, maps it to a vector space, and converts it into an item embedding representation v j Convert to. In this embodiment, the user embedded expression u obtained in step S2 kj,1 ,u kj,2 ,u kj,3 is the item embedding representation v obtained in step S2, which includes the above item purchase frequency. j will contain at least the price information of the target item.

[0056] In step S3, the weighting unit 13 of the group recommendation device 1 weights the user embedded expression u generated by the encoding unit 12 in step S2. kj,1 ,u kj,2 ,u kj,3 and the item embedding representation v j is input to the user identification machine learning model 31, and the user identification machine learning model 31 derives an attention weight for each user. The weighting unit 13 also applies the attention weight for each user output from the user identification machine learning model 31 to the corresponding user embedding representation u kj,1 ,u kj,2 ,ukj,3 and convert it into a weighted user embedding representation. Details of the weighting process performed using this user identification machine learning model 31 will be described later with reference to Fig. 5. Furthermore, the user identification machine learning model 31 may be a machine learning model equipped with a pre-trained neural attention mechanism, and details of the pre-training will be described later with reference to Fig. 9.

[0057] In step S4, the aggregator 14 of the group recommendation device 1 aggregates the weighted user-embedded expressions for the users generated by the weighter 13 in step S3. In step S5, the aggregator 14 of the group recommendation device 1 also aggregates the weighted user embedding expressions to generate a group embedding expression g l (j) is generated. The group embedding representation g generated in step S5 l (j) is a group embedding representation that represents the preferences of the user group to which the target user belongs.

[0058] The processes from steps S1 to S5 may be performed by the group recommendation device 1 using the machine learning model 32 for generating group embedded expressions and the machine learning model 31 for user identification, or by implementing these machine learning models 31 and 32. Alternatively, the machine learning model 31 for user identification may be incorporated as a module into the machine learning model 32 for generating group embedded expressions. Here, the machine learning model broadly includes networks, algorithms, modules, etc. that can learn parameters through machine learning.

[0059] In step S6, the group preference prediction unit 15 of the group recommendation device 1 calculates the group embedding representation g of the user group generated in step S5. l (j) and the item embedding representation of the target item v jis propagated through a multi-layer network, for example, a multi-layer network such as a Multi Layer Perceptron (MLP). In step S7, the group preference prediction unit 15 of the group recommendation device 1 calculates the group embedding representation g of the target user group. l (j) and the item embedding representation of the target item v j The result of processing this data with a multi-layer network is used to predict the score for the target item for the target user group.

[0060] The group preference prediction unit 15 calculates the group embedding representation g of the target user group output by the machine learning model 32 for generating group embedding representations. l (j) and the item embedding representation of the target item v j The group embedding representation g to be input to the machine learning model for item recommendation 33 l (j) and item embedding representation v j may be input to the machine learning model for item recommendation 33, causing the machine learning model for item recommendation 33 to predict the score for the target item of the target user group. The score predicted in step S6 is an index that quantitatively indicates the possibility that a target item will be purchased when recommended to a target user group. In other words, an item with a higher score predicted in step S6 can be evaluated as an item that can be expected to have a higher advertising effect when recommended.

[0061] The group recommendation device 1 may repeat the processes from steps S1 to S7 multiple times until a score for the desired item is obtained. In step S8, the recommended item output unit 16 of the group recommendation device 1 determines, based on the scores of the target items output by the group preference prediction unit 15, items for which a score equal to or greater than a predetermined threshold or a learnable threshold is calculated as items to be recommended, and outputs the items to be recommended or a list of multiple items to be recommended to a device such as a display device of a user belonging to the target user group.

[0062] <Detailed procedure for weighting user-embedded expressions> FIG. 5 is a flowchart showing an example of detailed processing procedures of the weighting process executed by the group recommendation device 1 in step S3 of FIG. In step S2 of FIG. 4, the user embedding representation u of the user belonging to the target user group is calculated. kj,1 ,u kj,2 ,u kj,3 and the item embedding representation of the target item v j When the user embedding expression u is generated, in step S31 of FIG. 5, the weighting unit 13 of the group recommendation device 1 receives the encoded user embedding expression u from the encoding unit 12. kj,1 ,u kj,2 ,u kj,3 In this embodiment, the user-embedded expression u obtained by the weighting unit 13 in S31 is obtained. kj,1 ,u kj,2 ,u kj,3 includes at least a feature that encodes the item purchase frequency. In step S32, the weighting unit 13 of the group recommendation device 1 receives the encoded item embedding representation v from the encoding unit 12. j In this embodiment, the item embedding representation v obtained by the weighting unit in S32 is obtained. j contains at least a feature that encodes the price of the target item.

[0063] In this embodiment, the attention weight for each user is derived based on the correlation between the price of the target item and the item purchase frequency. As described above, the item purchase frequency refers to the frequency with which the user has previously purchased an item on the e-commerce site (or e-commerce platform) being predicted, and includes the number of times the user has purchased not only the target item but also multiple items, including other items available on the e-commerce site. However, if the user or user group has had many past interactions with the target item being predicted, i.e., if the target item has been purchased many times in the past, the purchase frequency may be limited to the purchase frequency of only the target item or items similar to the target item among all items available on the e-commerce site.

[0064] Frequent buyers, users who purchase items more frequently than other users in a user group, are considered to be critical users who have more influence over the decision to purchase items within the user group. However, the inventors of this application found that this tendency changes depending on the price of the item. Here, a user with a high item purchase frequency (frequent buyer) may be determined relative to the item purchase frequency of other users in the user group. For example, a user with an item purchase frequency that is more than twice the average item purchase frequency of users in the same user group may be determined to be a user with a high item purchase frequency (frequent buyer). Alternatively, a threshold may be set for the minimum number of past purchases on the e-commerce site to be predicted, and a user with a number of item purchases exceeding this threshold may be determined to be a user with a high item purchase frequency.

[0065] Specifically, in this embodiment, the attention weight for each user is derived based on the following finding: when the price of the target item is relatively low, users who frequently purchase the item become critical users who have a greater influence on the decision to purchase the target item within the user group; on the other hand, when the price of the target item is relatively high, users who frequently purchase the target item lose their influence on the decision to purchase the target item within the user group. This finding was newly discovered and verified by the inventors of the present application by analyzing the users who had the greatest influence on the results of item recommendations to a user group.

[0066] Based on the above findings, the group recommendation device 1 according to this embodiment trains the user identification machine learning model 31 to model the process of deriving a user's attention weight from the price of an item and the user's interaction frequency, such as the purchase frequency of the item. Specifically, the user identification machine learning model 31 is a machine learning model trained to assign a higher weight to users who purchase an item more frequently when the price of the item is low, compared to when the price of the item is high. To infer the preferences of a user group for items using the thus trained user identification machine learning model 31, the group recommendation device 1 according to this embodiment evaluates users who frequently purchase the item as more important users when the price of the target item is relatively low, and derives a larger attention weight so that the preferences of these users have a more dominant influence on the inference of the preferences of the user group for items. On the other hand, when the price of the target item is relatively high, the group recommendation device 1 according to this embodiment derives a smaller attention weight so that the influence of the preferences of users who frequently purchase the item on the inference of the preferences of the user group for items is reduced.

[0067] By applying the user identification machine learning model 31 configured in this manner, according to this embodiment, important users can be determined more selectively according to the attributes of items, rather than simply selecting popular items as recommended items, thereby improving the accuracy of inferring the preferences of a user group and recommending items. Furthermore, there is no need to train the user identification machine learning model 31 by contrasting user information of non-member users who do not belong to the target user group with user information of member users who belong to the target group. This makes it possible to infer the preferences of a user group and recommend items with higher accuracy without preparing a large-scale learning dataset.

[0068] Returning to FIG. 5, in step S33, the weighting unit 13 of the group recommendation device 1 calculates the user embedding expression u acquired in S31. kj,1 ,u kj,2 ,u kj,3 and the item embedding representation v obtained in S32 j is input to the machine learning model for user identification 31, and the machine learning model for user identification 31 is caused to derive the attention weight to be assigned to the user-embedded expression of the target user. In this embodiment, the user identification machine learning model 31 may derive the attention weight using, for example, an adaptive sigmoid function (σ) shown in the following equation 1.

[0069]

number

[0070] where W tо is the attention weight of user t for item о within the user group, α о is inversely proportional to the price of item о, p t and , respectively, represent the proportion to the purchase frequency of user t. 1 can be any scalar. Figure 6 shows the adaptive sigmoid function in Equation 1, where different values ​​are substituted for the price variable of item о, and the corresponding attention weight W tо 6 is a diagram showing an example of calculation of the user identification machine learning model 31. As shown in FIG. 6, the user identification machine learning model 31 models the relationship between the price of an item, the interaction between the user and the item (indicated by, for example, the frequency of purchase of the item), and the attention weight for the user. Here, the information on the purchase frequency of an item may be normalized to a value between 0 and 5, for example. The price information may include the regular price, sale price, whether or not the item is on sale, the price range class, etc. The price information may be normalized to a value between 0.01 and 1, for example.

[0071] In Figure 6, the inverse proportional value α for the price of item o о The corresponding attention weights W are 0.1, 0.5, 1, and 1.5, respectively. tо are plotted as solid, dashed, dashed and dotted lines, respectively. о Since is inversely proportional to item о, α shown by the solid line о = 0.1, the price of item о is the highest, and α о = 1.5 is the lowest price for item o. As can be seen from Figure 6, the price of item о is the highest α о When σ = 0.1, the attention weight W is shown by the solid line. tо The value of increases linearly slightly as the purchase frequency of item о of user t increases, but does not change significantly. о When = 1.5, the attention weight W tо The value of is low when the purchase frequency of item o by user t is low, and increases significantly to a high value as the purchase frequency increases. That is, as the price of item о decreases, α о becomes higher, and the output W of the sigmoid function in Eq. tоshows a steep slope. This means that users who purchase items frequently (frequent buyers) have significantly greater attention than users who purchase items infrequently. On the other hand, if the price of item о is high, α о becomes lower, and the output W of the sigmoid function in Eq. tо The slope of decreases, which means that the difference between users who purchase items frequently (frequent buyers) and users who purchase items infrequently (non-frequent buyers) decreases.

[0072] In this manner, in this embodiment, for items with relatively low prices, the attention weight W tо The sigmoid function is adapted so that the value of changes significantly in response to changes in the user's item purchase frequency. Note that the adaptive sigmoid function shown in Equation 1 is just an example, and is not limited to this function that can be implemented in the user identification machine learning model 3. Other activation functions, such as a softmax function, may be used as long as they can model the relationship between the price of an item, the interaction between the user and the item (e.g., indicated by purchase frequency), and the attention weight for the user.

[0073] In this embodiment, the adaptive sigmoid function is implemented in the user identification machine learning model 31, which is a neural attention mechanism, to dynamically derive attention weights to be assigned to user embedded expressions for each user belonging to a user group. These attention weights can be tuned by having the user identification machine learning model 31 pre-train the relationship between the item price, the user's interaction frequency with the item, and the attention weights to be assigned to users within the user group. The pre-training unit 17 causes the user identification machine learning model 31 to learn attention weights that indicate the influence or importance of each user in the user group's decision-making to purchase a target item. By using the machine learning model 31 for user identification configured as described above, the group recommendation device 1 according to this embodiment can derive attention weights with high accuracy, a simple configuration, and low load.

[0074] <Processing Procedure Using the Machine Learning Model 32 for Generating Group Embedding Representations and the Machine Learning Model 33 for Recommending Items in the Present Embodiment> FIG. 7 is a conceptual diagram showing an example of a processing procedure when the process of deriving preferences of a user group executed by the group recommendation device 1 according to this embodiment is implemented in the machine learning model 32 for generating group embedded expressions. FIG. 8 is a conceptual diagram showing an example of a processing procedure when the processing of recommending items to a user group executed by the group recommendation device 1 according to this embodiment is implemented in the machine learning model 33 for item recommendation. The basic network configuration of the machine learning model 32 for generating group embedded expressions is the same as the first network configuration shown in Fig. 2, and the basic network configuration of the machine learning model 33 for recommending items is the same as the second network configuration shown in Fig. 3. However, in Fig. 7, the neural attention mechanism 204 in Fig. 2 is replaced with the machine learning model 31 for user identification according to this embodiment.

[0075] 7, user behavior history information of multiple users belonging to the same user group 701 is input to the machine learning model 32 for generating group embedded expressions. The input user behavior history information is converted into user embedded expressions u for each user in the embedding layer 202 of the machine learning model 32 for generating group embedded expressions. p , u q is encoded as The machine learning model for generating group embedding representations 32 also receives as input item information 22 of the target item for which the preferences of the user group are to be inferred. The input item information 22 includes an item identifier from which at least the price information of the target item can be derived. Such item information 22 is used to generate item embedding representations i in the embedding layer 202 of the machine learning model for generating group embedding representations 32. о may be encoded as User embedding expression u p , u q and item embedding representation i о is input to a machine learning model for user identification 31, which acts as a neural attention mechanism. Here, the user embedding representation u p , u q The item embedding representation i о The price information of the target item o is encoded in

[0076] The user identification machine learning model 31 implements the adaptive sigmoid function (σ) in the above formula 1 and generates the user embedding representation u p , u q and item embedding representation i о Applying an adaptive sigmoid function (σ) to the attention weights W corresponding to each user p and q p , W q are derived respectively. Attention weights W for each user p and q derived by the user identification machine learning model 31 p , W q is a user embedding representation u in the pooling layer 203 of the machine learning model 32 for generating group embedding representations. p , u q are applied to generate weighted user embedded expressions 702 and 703, respectively. The weighted user embeddings 702, 703 are pooled in the pooling layer 203 of the machine learning model 32 to generate a single group embedding g lоand output from the output layer 205. Specifically, the group embedding representation g lо is derived by the following equation 2.

[0077]

number

[0078] where g lо is the group embedding obtained as the sum of the weighted user embeddings of users belonging to user group l. W tо is the attention weight for the target item о, and u t is the user embedding representation of user t. This group embedding representation g lо is an embedding representation specific to the target item o.

[0079] 8, the representation layer 301 of the item recommendation machine learning model 33 includes the group embedding representation g output by the group embedding representation generation machine learning model 32 shown in FIG. lо and the item embedding representation of the target item i о is entered. The representation layer 301 of the machine learning model for item recommendation 33 also includes a group embedding representation b specific to a user group l. l This group embedding representation b l is the user embedding representation u p , u q The group embedding representation g generated by aggregating lо In this case, the group embedding g generated by aggregating the user embeddings is lо The user group specific group embedding representation b l are added as shown in Equation 3 below to obtain the group embedding representation f lо may be calculated and fed to the subsequent pooling layer 802. flо =g lо +b l (Formula 3)

[0080] In this way, by adding group embedding expressions specific to user groups that are independent of items to group embedding expressions that aggregate user embedding expressions, it is possible to reflect more specific attributes of each user group, such as age segments, segments of residence or workplace, segments of annual income or hobbies, etc., in the score calculation. In the pooling layer 802 of the machine learning model 33 for item recommendation, the group embedding representation f lо and the item embedding representation i of the target item о о are concatenated to generate a group-item concatenated embedded representation. lо may be concatenated as in Equation 4 below.

[0081]

number

[0082] In Equation 4, the group-item concatenated embedding representation e lо is a set of three embeddings, namely, the group embedding f lо , item embedding representation i о , and an embedding representation that concatenates the element-wise product or Hadamard product of the group embedding representation and the item embedding representation. The element-wise product of the group embedding and the item embedding models the interaction between the two embeddings, and is therefore effective for predicting preferences for items in a group. However, the calculation results in a loss of original information. To compensate for this loss of information, in this embodiment, the group embedding f lо and item embedding representation i о are further concatenated for element-wise product.

[0083] Group-item concatenated embedding representation elо is supplied as a pooling vector to the subsequent multi-layer perception (MLP) 803, and is propagated through multiple layers of the MLP 803 to produce the score ^y lо This score ^y lо is a score that predicts the preference of user group l for item o, and is an indicator of the advertising effectiveness of recommending item o. This MLP 803 is a stack of multiple fully connected layers, and can obtain nonlinear and high-order correlations between users, user groups, and items.

[0084] Referring to FIG. 8, the machine learning model for item recommendation 33 also generates a core ^y that predicts the preferences of user group l for item o. lо At the same time, the score predicting each user's preference for item o may be output. In this case, similar to the case where the group embedding representation is input, the user embedding representation u of the target user p is p Let i be the item embedding representation of the target item. о The user embedding representation u is input to the representation layer 301 of the item recommendation machine learning model 33, and the pooling layer 802 p and the item embedding representation i о The user-item concatenated embeddings are generated by concatenating them in the same way as in Equation 4. The user-item concatenated embeddings generated here are composed of three embeddings, namely, the user embeddings u p , item embedding representation i о , and an embedding representation that concatenates the element-wise product or Hadamard product of the user embedding representation and the item embedding representation. The user-item embedded representation is fed as a pooling vector to the subsequent multi-layer perception (MLP) 803, and is propagated through multiple layers of the MLP 803 to produce a score ^x pо This score^x is output by MLP803. pо is a score that predicts user p's preference for item o, and serves as an indicator of the advertising effectiveness of recommending item o.

[0085] In this way, by sharing the model parameters contained in the same machine learning model for item recommendation 33 between the output of the score predicting the preferences of a user group and the output of the score predicting the preferences of a user belonging to that group, information on interactions between individual users and items is additionally applied to the machine learning model for item recommendation 33, making it possible to learn the group embedding representation of the user group with higher accuracy. The multilayer perceptron (MLP) shown in Figure 8 is a type of forward propagation neural network consisting of at least an input layer, a hidden layer (an intermediate layer), and an output layer. Each node in each layer, except for the input node, operates as a perceptron, inputting the sum of weighted inputs into a nonlinear activation function to obtain the output for the next layer. The machine learning model 33 for item recommendation shown in FIG. 8 can be implemented in, for example, Neural Collaborative Filtering (NCF), which learns interactions (interrelationships) between users and items based on the user's behavioral history regarding items.

[0086] <Pre-training of each machine learning model> FIG. 9 is a flowchart showing an example of detailed processing procedures for pre-learning each machine learning model according to this embodiment. Referring to FIG. 9, in step S91, the pre-learning unit 17 of the group recommendation device 1 prepares a training dataset for pre-learning each machine learning model, and stores the prepared training dataset in a storage device. Specifically, the learning dataset prepared for the user identification machine learning model 31 includes training data in which the frequency of user interactions with items (e.g., item purchase frequency) is labeled and the item prices included in the item information are labeled from user behavior history information, which is a behavior history for items included in the EC usage history 21, and the attention weights of users within a user group are annotated as ground truth data for pairs of user behavior history information and item information. The ground truth data for the user's attention weights may be obtained by backpropagation from the results of item recommendations, such as items that led to more purchases, greater advertising effectiveness, or were highly rated in user reviews. The learning dataset prepared for the machine learning model that integrates the machine learning model 32 for generating group embedded representations and the machine learning model 33 for recommending items may include training data in which evaluation values ​​(e.g., ratings obtained from user reviews) and evaluation ranks of recommended items are annotated as correct answer data for pairs of user behavior history information, which is a user's behavior history regarding items, and item information.

[0087] In step S92, the pre-learning unit 17 inputs the learning data set prepared in step S91 into each machine learning model, and trains each machine learning model. In step S93, the pre-learning unit 17 repeats the pre-learning of each machine learning model until the loss between the output output by each machine learning model in step S92 and the correct answer data becomes smaller and falls below a predetermined threshold of loss. Specifically, for example, when the machine learning models 31, 32, and 33 are integrated into one machine learning model, in order to evaluate the pre-learning and optimize the inference results, the machine learning model outputs a preference score ^y for the item о of the user group l. lо is applied to a predetermined loss function and learning is repeated until the loss becomes smaller. For example, if explicit feedback on scores and the item recommendation results based on them is not available, a regression pairwise loss function such as the one shown in Equation 5 below can be used.

number

number

[0088] As described above, according to this embodiment, the group recommendation device encodes a user's behavioral history for items, extracts the user's preferences for items for each user, derives a weight indicating the user's importance within a user group using information about the items for each user, weights the extracted user preferences with the derived weight, and aggregates the weighted preferences to generate preferences for a user group to which the user belongs. Here, the weight for each user is derived differently depending on the correlation between the target user's transactions for items and information about the items. The group recommendation device according to this embodiment further predicts a score indicating the user group's preferences for items from the generated user group preferences and information about the corresponding items.

[0089] This makes it possible to derive the preferences of a group to which a user belongs more appropriately and with higher accuracy from the preferences of the user items, even for a large user group with many users, with a simple configuration.As a result, it is possible to realize more effective item recommendations for a user group made up of a variety of users who use an e-commerce site.

[0090] <Hardware configuration of the group recommendation device> FIG. 10 is a diagram showing a non-limiting example of the hardware configuration of the group recommendation device 1 according to this embodiment. The group recommendation device 1 according to this embodiment can be implemented on any single or multiple computers, mobile devices, or any other processing platform. 10 shows an example in which the group recommendation device 1 is implemented in a single computer, but the group recommendation device 1 according to this embodiment may be implemented in a computer system including multiple computers, which may be connected to each other via a wired or wireless network so as to be able to communicate with each other.

[0091] 10, the group recommendation device 1 may include a CPU 101, a ROM 102, a RAM 103, an HDD 104, an input unit 105, a display unit 106, a communication I / F 107, and a system bus 108. The group recommendation device 1 may also include an external memory. The CPU (Central Processing Unit) 101 comprehensively controls the operations of the group recommendation device 1, and controls each component (102 to 107) via a system bus 108, which is a data transmission path. Note that instead of or in addition to the CPU 101, the group recommendation device 1 may be provided with a GPU (Graphics Processing Unit), which may be used to execute pre-learning and inference processing for the machine learning models 31 to 33.

[0092] The ROM (Read Only Memory) 102 is a nonvolatile memory that stores control programs and the like required for the CPU 101 to execute processing. Note that the programs may be stored in a nonvolatile memory such as an HDD (Hard Disk Drive) 114 or an SSD (Solid State Drive), or in an external memory such as a removable storage medium (not shown). RAM (Random Access Memory) 103 is a volatile memory that functions as the main memory, work area, etc. of CPU 101. That is, when executing a process, CPU 101 loads necessary programs, etc. from ROM 102 into RAM 103 and executes the programs, etc. to realize various functional operations.

[0093] The HDD 104 stores, for example, various data and information required when the CPU 101 performs processing using a program. The HDD 104 also stores, for example, various data and information obtained when the CPU 101 performs processing using a program. The input unit 105 is composed of a keyboard and a pointing device such as a mouse. The display unit 106 is configured with a monitor such as a liquid crystal display (LCD). The display unit 106 may provide a GUI (Graphical User Interface), which is a user interface for inputting various parameters used in machine learning processing, communication parameters used in communication with other devices, and the like, into the group recommendation device 1.

[0094] The communication I / F 107 is an interface that controls communication between the group recommendation device 1 and an external device. The communication I / F 107 provides an interface with a network and communicates with external devices via the network. Various data, parameters, and the like are transmitted and received between the external devices and the communication I / F 107. In this embodiment, the communication I / F 107 may communicate via a wired LAN (Local Area Network) or a dedicated line conforming to a communication standard such as Ethernet (registered trademark). However, the network available in this embodiment is not limited to this and may be configured as a wireless network. This wireless network includes wireless PANs (Personal Area Networks) such as Bluetooth (registered trademark), ZigBee (registered trademark), and UWB (Ultra Wide Band). It also includes wireless LANs (Local Area Networks) such as Wi-Fi (Wireless Fidelity) (registered trademark) and wireless MANs (Metropolitan Area Networks) such as WiMAX (registered trademark). It also includes wireless WANs (Wide Area Networks) such as LTE / 3G, 4G, and 5G. Note that the network may be any network that connects devices to each other and enables communication, and the communication standard, scale, and configuration are not limited to those described above.

[0095] At least some of the functions of each element of the group recommendation device 1 shown in Fig. 1 can be realized by the CPU 101 executing a program. However, at least some of the functions of each element of the group recommendation device 1 shown in Fig. 1 may be operated as dedicated hardware. In this case, the dedicated hardware operates under the control of the CPU 101.

[0096] Although specific embodiments have been described above, these embodiments are merely examples and are not intended to limit the scope of the present invention. The devices and methods described herein may be embodied in forms other than those described above. Furthermore, appropriate omissions, substitutions, and modifications may be made to the above-described embodiments without departing from the scope of the present invention. Such omissions, substitutions, and modifications are included within the scope of the claims and their equivalents, and belong to the technical scope of the present invention.

[0097] (Embodiments of the present disclosure) The present disclosure includes the following embodiments. [1] An information processing device comprising: an encoding unit that encodes a user's behavioral history regarding an item and generates a first embedded representation for each user; a weighting unit that derives a weight for each user based on information regarding the price of the item and the behavioral history and weights the first embedded representation generated by the encoding unit with the derived weight; and an aggregation unit that aggregates the first embedded representations weighted by the weighting unit and generates, for a group to which the user belongs, a second embedded representation that indicates the group's preference for the item.

[0098] [2] The information processing device described in [1], wherein the weighting unit derives a different weight for the user depending on the price of the item.

[0099] [3] The information processing device according to [1] or [2], characterized in that the weighting unit derives the weight so that the lower the price of the item, the larger the weight.

[0100] [4] The information processing device described in any one of [1] to [3], characterized in that the weighting unit derives the weight based on information regarding the price of the item and the user's purchase frequency of the item on an e-commerce site, which is encoded in the first embedded representation.

[0101] [5] The information processing device described in any one of [1] to [4], characterized in that the encoding unit further encodes information including the price of the item to generate a third embedded representation, and the weighting unit derives the weight based on information regarding the price of the item encoded in the third embedded representation.

[0102] [6] The information processing device described in any one of [1] to [5], characterized in that the weighting unit derives the weights by applying information about the price of the item and a first machine learning model that has learned the behavioral history to the first embedded representation.

[0103] [7] The information processing device described in [6], characterized in that the first machine learning model is trained to output the weight so that the lower the price of the item, the larger the weight.

[0104] [8] The information processing device according to [6] or [7], characterized in that the first machine learning model implements an activation function with the price of the item and the user's purchase frequency of the item on an e-commerce site as variables to derive the weights.

[0105] [9] The information processing device according to any one of [1] to [8], further comprising a prediction unit that predicts a score for the item in the group by concatenating the second embedded representation of the group and a third embedded representation that encodes information including the price of the item, and propagating the concatenated embedded representation through a multi-layered neural network.

[0106]

[10] The information processing device described in [9], characterized in that the prediction unit further concatenates the second embedded representation and the third embedded representation to the concatenated embedded representation, and propagates the concatenated embedded representation to the multiple-layer neural network.

[0107]

[11] The information processing device described in [9] or

[10] , characterized in that the prediction unit adds a fourth embedded representation that encodes an attribute unique to the group to the second embedded representation of the group, and concatenates the second embedded representation to which the fourth embedded representation has been added to the third embedded representation.

[0108]

[12] The information processing device described in any one of [9] to

[11] , characterized in that the prediction unit further predicts the score for the item of the user simultaneously with the score for the item of the group by concatenating the third embedded representation of the item and the first embedded representation of the user and propagating the concatenated embedded representation through a multi-layered neural network.

[0109]

[13] The information processing device described in any one of [9] to

[12] , characterized in that the prediction unit uses the multi-layered neural network to predict the score for the item of the group and the score for the item of the user.

[0110]

[14] An information processing method executed by an information processing device, comprising the steps of: encoding a user's behavioral history regarding an item to generate a first embedded representation for each user; deriving a weight for each user based on information regarding the price of the item and the behavioral history, and weighting the generated first embedded representation with the derived weight; and aggregating the weighted first embedded representations to generate a second embedded representation for a group to which the user belongs, the second embedded representation indicating the group's preference for the item.

[0111]

[15] A computer-readable recording medium storing an information processing program for causing a computer to execute information processing, the program causing the computer to execute processes including: an encoding process for encoding a user's behavioral history for an item to generate a first embedded representation for each user; a weighting process for deriving a weight for each user based on information about the price of the item and the behavioral history, and weighting the first embedded representation generated by the encoding process with the derived weight; and an aggregation process for aggregating the first embedded representations weighted by the weighting process to generate, for a group to which the user belongs, a second embedded representation that indicates the group's preference for the item. [Explanation of symbols]

[0112] 1...Group recommendation device, 2...Transaction information storage device, 3...Machine learning model storage device, 11...Input unit, 12...Encoding unit, 13...Weighting unit, 14...Aggregation unit, 15...Group preference prediction unit, 16...Recommended item output unit, 17...Pre-learning unit, 21...EC site usage history, 22...Item information, 31...Machine learning model for user identification, 32...Machine learning model for generating group embedded representations, 33...Machine learning model for item recommendation, 101...CPU, 102...ROM, 103...RAM, 104...HDD, 105...Input unit, 106...Display unit, 107...Communication I / F, 108...System bus

Claims

1. an encoding unit that encodes a behavior history of a user with respect to an item to generate a first embedded representation for each user; a weighting unit that derives a weight for each user based on information about the price of the item and the behavioral history, and weights the first embedded representation generated by the encoding unit with the derived weight; an aggregation unit that aggregates the first embedded representations weighted by the weighting unit to generate, for a group to which the user belongs, a second embedded representation that indicates a preference of the group for the item.

1. An information processing device comprising:

2. The weighting unit derives different weights for the user depending on the price of the item.

2. The information processing apparatus according to claim 1, wherein:

3. The weight assigning unit derives the weight such that the lower the price of the item, the greater the weight.

2. The information processing apparatus according to claim 1, wherein:

4. the weighting unit derives the weight based on information about the price of the item and the user's purchase frequency of the item on an e-commerce site, the purchase frequency being encoded in the first embedded representation; 2. The information processing apparatus according to claim 1, wherein:

5. The encoding unit further encodes information including a price of the item to generate a third embedded representation; the weighting unit derives the weight based on information about the price of the item encoded in the third embedded representation.

2. The information processing apparatus according to claim 1, wherein:

6. the weighting unit derives the weights by applying a first machine learning model, which has been trained on information about the prices of the items and the behavioral history, to the first embedded representation; 2. The information processing apparatus according to claim 1, wherein:

7. The first machine learning model is trained to output the weights such that the lower the price of the item, the larger the weights.

7. The information processing apparatus according to claim 6,

8. The first machine learning model derives the weights by implementing an activation function whose variables are the price of the item and the user's purchase frequency of the item on an e-commerce site.

7. The information processing apparatus according to claim 6,

9. The information processing device further comprises: a prediction unit that predicts a score for the item in the group by concatenating the second embedded representation of the group and a third embedded representation that encodes information including the price of the item, and propagating the concatenated embedded representation through a multi-layered neural network; 2. The information processing apparatus according to claim 1, wherein:

10. the prediction unit further concatenates the second embedding representation and the third embedding representation to the concatenated embedding representation, and propagates the concatenated embedding representation to the multi-layered neural network.

10. The information processing apparatus according to claim 9,

11. the prediction unit adds a fourth embedded representation in which an attribute specific to the group is encoded to the second embedded representation of the group, and concatenates the second embedded representation to which the fourth embedded representation has been added to the third embedded representation.

10. The information processing apparatus according to claim 9,

12. The prediction unit further predicts the score for the item of the user simultaneously with the score for the item of the group by concatenating the third embedded representation of the item and the first embedded representation of the user and propagating the concatenated embedded representation through a multi-layered neural network.

10. The information processing apparatus according to claim 9,

13. the prediction unit uses the multiple layer neural network in common for predicting scores for the items of the group and predicting scores for the items of the user; 10. The information processing apparatus according to claim 9,

14. An information processing method executed by an information processing device, encoding a user's behavior history with respect to items to generate a first embedding representation for each user; deriving a weight for each user based on information about the price of the item and the behavioral history, and weighting the generated first embedding representation with the derived weight; aggregating the weighted first embeddings to generate, for a group to which the user belongs, a second embedding that indicates the group's preference for the item. An information processing method comprising:

15. A computer-readable recording medium storing an information processing program for causing a computer to execute information processing, the program including: an encoding process for encoding a behavior history of a user with respect to an item to generate a first embedded representation for each user; a weighting process of deriving a weight for each user based on information about the price of the item and the behavior history, and weighting the first embedded representation generated by the encoding process with the derived weight; an aggregation process for aggregating the first embedded representations weighted by the weighting process to generate a second embedded representation indicating a preference of the group for the item for a group to which the user belongs; A computer-readable recording medium for executing a process including:

Images