Information processing device, method, and program

Abstract

To generate a trained model with improved interpretability.SOLUTION: An information processing device according to an embodiment includes an acquisition unit, a persona generation unit, an extraction unit, a recommendation loss calculation unit, a comparison loss calculation unit, and an update unit. The acquisition unit acquires graph data. The persona generation unit generates additional graph data to which one or more personas have been added as nodes. The extraction unit uses a model to extract feature quantities for each of the subject and persona from the additional graph data. The recommendation loss calculation unit uses each feature quantity to calculate a first loss corresponding to the difference between the similarity between nodes connected by edges and the similarity between nodes not connected by edges. The comparison loss calculation unit uses each feature quantity to calculate a second loss corresponding to the difference between the similarity between the nodes connected to the subject's node and the similarity between the nodes connected to the persona's node. The update unit updates the model using the first loss and the second loss.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] FIELD Embodiments of the present invention relate to an information processing device, method, and program. [Background technology]

[0002] User product purchase data and review data can be treated as a bipartite graph, where users and items are represented as nodes and purchase relationships and review relationships are represented as edges. In recent years, recommendation system technology that uses graph analysis techniques such as graph neural networks (GNNs) on such bipartite graphs to recommend items to users has been attracting attention. Recommendation system technology that uses graph analysis incorporates high-level relationships between users and items, extracts features for each user and item, and trains a model so that the purchase probability between users and items with a purchasing relationship is greater than the purchase probability between users and items with no purchasing relationship.

[0003] A user can be divided into categories based on preferences, and each aspect of the user can be called a persona. By training a recommendation system on a persona, it is possible to expect recommendations that are more tailored to individual preferences. However, recommendation system technology learns the relative purchase probability (ranking) between a user / persona and an item they have purchased, and between a user / persona and an item they have not purchased. This means that it is unable to learn the relative relationship between which user or persona is more likely to purchase an item, resulting in poor interpretability. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] U.S. Patent No. 11,100,688 Summary of the Invention [Problem to be solved by the invention]

[0005] The present disclosure has been made to solve the above-mentioned problems, and aims to provide an information processing device, method, and program that can generate a trained model with improved interpretability. [Means for solving the problem]

[0006] The information processing device according to this embodiment includes an acquisition unit, a persona generation unit, an extraction unit, a recommendation loss calculation unit, a comparison loss calculation unit, and an update unit. The acquisition unit acquires graph data in which multiple entities are represented by nodes for each of two or more categories and relationships between the entities are represented by edges connecting the nodes. The persona generation unit generates additional graph data to which one or more personas having representative characteristics of the entities are added as nodes. The extraction unit extracts entity features and persona features from the additional graph data using a feature extraction model. The recommendation loss calculation unit uses the entity features and the persona features to calculate a first loss corresponding to the difference between the similarity between nodes connected by the edges and the similarity between nodes not connected by the edges. The comparison loss calculation unit uses the entity features and the persona features to calculate a second loss corresponding to the difference between the similarity between the entity's node and a node connected to the entity's node and the similarity between the persona's node and a node connected to the persona's node. An update unit updates the model using the first loss and the second loss. [Brief explanation of the drawings]

[0007] [Figure 1] FIG. 1 is a block diagram showing an information processing apparatus according to an embodiment of the present invention. [Figure 2] 10 is a flowchart showing an example of the operation of the information processing device according to the present embodiment. [Figure 3] FIG. 10 is a diagram showing an example of additional graph data according to the embodiment. [Figure 4] FIG. 10 is a diagram showing an example of additional graph data when a community is formed. [Figure 5] FIG. 10 is a diagram for explaining the relationship between nodes when calculating the recommendation loss based on additional graph data. [Figure 6] FIG. 10 is a diagram for explaining the relationship between nodes when calculating the recommendation loss based on additional graph data. [Figure 7] FIG. 10 is a diagram for explaining the relationship between nodes when calculating a comparison loss based on additional graph data. [Figure 8] FIG. 1 is a diagram showing an example of a hardware configuration of an information processing apparatus. DETAILED DESCRIPTION OF THE INVENTION

[0008] Hereinafter, an information processing device, method, and program according to the present embodiment will be described in detail with reference to the drawings. In the following embodiments, parts with the same reference numerals perform similar operations, and redundant explanations will be omitted as appropriate. An information processing device 10 according to this embodiment will be described with reference to the block diagram of FIG. The information processing device 10 according to this embodiment includes a storage unit 101 , an acquisition unit 102 , a persona generation unit 103 , an extraction unit 104 , a recommendation loss calculation unit 105 , a comparison loss calculation unit 106 , and an update unit 107 .

[0009] The storage unit 101 stores graph data, models, and the like. Graph data is data related to a graph in which, for each category, which is a collection of multiple entities, each entity is represented by a node and the relationships between the entities are represented by edges connecting the nodes. In this embodiment, an entity represented by a node in a graph is referred to as an entity node. Graph data is, for example, a bipartite graph having two categories, and is data including relationships between the two categories, in which an entity node belonging to one category is connected by an edge to an entity node belonging to the other category. Specifically, in graph data showing a purchasing relationship between a person and a product (including a service), the person and the product are each a category, individuals belonging to the "person" category and individual products belonging to the "product" category are each represented by a node, and the purchasing relationship is represented by an edge. In addition, in graph data showing a visiting relationship between a person and a place such as a tourist spot or a commercial facility, the person and the place are each a category, individuals belonging to the "person" category and names of tourist spots and commercial facilities belonging to the "place" category are each represented by a node, and the visiting relationship is represented by an edge. A knowledge graph may be constructed by adding new entities (nodes) to each node according to the attributes of the node. For the sake of convenience, the following description will be given of two categories, "users" and "products," each of which includes multiple entities, but the graph data may also show the relationships between two or more categories.

[0010] The acquisition unit 102 acquires graph data from the storage unit 101 or from an external source. The persona generation unit 103 generates additional graph data in which one or more personas indicating the characteristics of the subject are added as nodes. A persona is information indicating characteristic aspects of the subject, such as the subject's preferences, to which the persona belongs.

[0011] The extraction unit 104 extracts subject features and persona features from the additional graph data using a model for extracting features. The recommendation loss calculation unit 105 uses the subject features and persona features to calculate a recommendation loss (also called the first loss) according to the difference between the similarity between nodes connected by edges and the similarity between nodes not connected by edges.

[0012] The comparison loss calculation unit 106 uses the subject features and persona features to calculate a comparison loss (second loss) according to the difference between the similarity between the subject node and the node connected to the subject node and the similarity between the persona node, which is a node related to the persona, and the node connected to the persona node. The update unit 107 updates the model used in the extraction unit 104 using the recommendation loss and the comparison loss.

[0013] Next, an example of the operation of the information processing device 10 according to this embodiment will be described with reference to the flowchart of FIG. In step SA1, the acquiring unit 102 acquires graph data. In the graph data, each node may have edges between nodes of the same subject. The initial value of each node may be attribute information such as a label, category, or parameter, or may be a degree in graph theory, an embedding vector obtained by Deepwalk or node2vec, or random initialization.

[0014] In step SA2, persona generation unit 103 generates a persona for each node based on the node information of the graph data and the edge information between the nodes, and adds the generated persona to the graph data as a node to generate additional graph data. Persona generation can be performed by persona generation unit 103 using, for example, a method called "SPLITTER" disclosed in Patent Document 1, an egocentric network, a graph Laplacian, or Ricci curvature, or any other commonly used method. Furthermore, persona generation unit 103 adds the edges duplicated from each node to the persona. It is assumed that the addition of personas is performed for all subject nodes included in at least one category of the graph data, but it may be performed for at least one category instead of all categories.

[0015] In step SA3, extraction unit 104 extracts subject features and persona features from the additional graph data obtained by the processing of step SA2. The subject features and persona features can be extracted from the additional graph data using, for example, a graph convolutional network. However, the subject features and persona features may also be extracted using other graph neural networks with message passing mechanisms, such as graph attention network (GAT), GraphSAGE, or graph isomorphism network (GIN).

[0016] In step SA4, the recommendation loss calculation unit 105 uses the subject feature amount and the persona feature amount to calculate a recommendation loss according to the difference between the similarity between adjacent nodes and the similarity between non-adjacent nodes. The nodes include subject nodes and persona nodes. The recommendation loss L in emphasis filtering is CF can be expressed by the following equation (1), for example.

[0017]

number

[0018] Here, u is the feature of the subject node related to the first subject (user). i is the feature of the node related to the second subject (product) to which the edge connects with the node related to the first subject (user). u p is the feature related to the persona node of the first entity (user). j is the feature related to the node related to the second entity (product) that is not connected by an edge to the node related to the first entity (user). y(u,i) is the similarity of the feature between the node related to the first entity and the node of the second entity that is connected by an edge, that is, the similarity of the feature between the user and the product or product persona that are adjacent to each other. Also, y(u p,j) is the similarity of the features between the persona node related to the persona of the first entity and the node of the second entity to which no edge is connected, that is, the similarity of the features between a user and a product or product persona that are not in an adjacent relationship.

[0019] In addition, possible adjacent relationships include a connection relationship via an edge between the node of the first subject and the node of the second subject, a connection relationship via an edge between the node of the persona of the first subject and the node of the second subject, a connection relationship via an edge between the node of the persona of the first subject and the node of the persona of the second subject, and a connection relationship via an edge between the node of the persona of the first subject and the node of the persona of the second subject. The similarity can be calculated by using the inner product of the subject feature and the persona feature, but any method that can calculate the similarity between features, such as cosine similarity, can also be used. CF Although it is assumed that the difference between similarities is used, any method can be used to calculate the loss. Here, the recommendation loss calculation unit 105 calculates the loss for all of the subjects (users or products) on one side of the graph data and the persona of the subject, and calculates the average of the losses as the recommendation loss L CF Output as

[0020] In step SA5, the comparison loss calculation unit 106 uses the subject feature and the persona feature to calculate the comparison loss L according to the difference between the similarity between the feature related to the subject node and the feature related to the adjacent node and the similarity between the feature related to the persona node and the feature related to the adjacent node. PS For example, it can be expressed as the following equation (2).

[0021]

number

[0022] The similarity is the recommendation loss L CFSimilarly, inner products, cosine similarities, etc. may be calculated. The difference in similarity may be the difference between similarities. The comparison loss calculation unit 106 calculates the loss for all of the features (users or products) of one subject node of the graph data and the persona of the subject, and calculates the average of the losses as the comparison loss L PS Output as In step SA6, the update unit 107 updates the recommendation loss L CF and comparative loss L PS Specifically, the model used in the extraction unit 104 is trained according to the recommendation loss L CF and comparative loss L PS From the total loss L TOT Calculate the total loss L TOT is expressed by equation (3).

[0023]

number

[0024] The update unit 107 updates the total loss L TOT Based on this, the recommendation loss L CF and comparative loss L PS The model parameters (weights and biases) are updated so that the difference in similarity between the nodes increases. The initial values ​​or features for the nodes may be used as trainable weights.

[0025] In step SA7, the update unit 107 determines whether the parameter update of the model is complete, that is, whether the training of the model is complete. The completion of the parameter update may be determined, for example, when the loss value becomes equal to or less than a threshold value. Alternatively, the training may be determined to be complete when the parameters have been updated a predetermined number of times. Alternatively, the training may be determined to be complete when the absolute value or the sum of the absolute values ​​of the parameter update amounts becomes a constant value. Note that the determination of whether the training is complete is not limited to the above example, and a termination condition generally adopted in machine learning may also be used. If the training is complete, the process ends. If the training is not complete, the process returns to step SA1 and continues processing on the next graph data.

[0026] Next, an example of added graph data when a persona according to this embodiment is added will be described with reference to FIG. FIG. 3 is a bipartite graph showing two different categories, A30 and B35, and their relationships. Category A30 is a set of users as subjects, and user information is represented by multiple subject nodes 31. Specifically, assume that eight subject nodes 31, from "User A" to "User H," belong to category A30. For ease of explanation, the node related to user A will also be referred to as subject node A, and subject nodes related to other users will also be referred to in the same way.

[0027] Moreover, category B35 is a set whose subject is a product, and the product is represented by a plurality of subject nodes 31. Specifically, assume that five subject nodes 31, from "product a" to "product e", belong to category B35. For ease of explanation, the node related to product a will also be called node a, and nodes related to other products will also be called similarly.

[0028] Furthermore, when a subject belonging to category A has a relationship with a subject belonging to category B, they are connected by an edge 32. A connection by an edge 32 can also be expressed as "nodes being adjacent." Specifically, subject node C of "user C" belonging to category A30 is connected by an edge 32 to subject node a of "product a," subject node b of "product b," subject node c of "product c," and subject node d of "product d," respectively, indicating that they have a relationship. In other words, subject node C is adjacent to subject node a, subject node b, subject node c, and subject node d.

[0029] In Figure 3, personas are generated for user A, user C, and product d, and persona node 33, which is a node corresponding to the persona, is added. Specifically, persona node A1, persona nodes C1, and C2 are generated for category A30, and persona nodes d1 and d2 are generated for category B35, and edge 32 is duplicated. For example, persona node A1 is connected to subject node a and subject node b by edge 32, respectively.

[0030] A community may be created in each entity. An example of additional graph data when a community is created is shown in FIG. A persona is assigned to each of the multiple communities to which a node belongs. In the example of Fig. 4, persona node A1 is added to community 40 from a subject node belonging to community 40, and persona node F1 is added to community 45 from a subject node belonging to community 45. Note that a group of nodes of the same subject that share a node of another subject may also be used as a community.

[0031] Next, the relationship between nodes when calculating the recommendation loss based on the additional graph data will be described with reference to FIGS. 5 and 6. FIG. In FIG. 5, subject node C is used as a reference and adjacent subject nodes or persona nodes are indicated by thick lines.

[0032] First, for each subject node and each persona node, the recommendation loss calculation unit 105 calculates a first similarity of feature amounts between a subject node or persona node belonging to category A30 and a subject node or persona node belonging to category B35 that is connected by edge 32.

[0033] As shown in FIG. 5, the first similarity for a subject node C is calculated from the similarity between the subject features of user C corresponding to the subject node C and the subject features of product a corresponding to subject node a, product b corresponding to subject node b, product c corresponding to subject node c, and product d corresponding to subject node d, which are adjacent to the subject node C, or the similarity between the subject features and persona features corresponding to persona node d1.

[0034] In FIG. 6, subject nodes or persona nodes that are not adjacent to persona node C2, which is the persona of subject node C, are indicated by thick dashed lines. The recommendation loss calculation unit 105 calculates a second similarity of features between a persona node belonging to category A30 and a subject node or persona node belonging to category B35 that is not connected by edge 32. As shown in Fig. 6 , the second similarity for persona node C2 is calculated from either the similarity between the persona features of persona C2 of user C and the subject features of products a, b, and e that correspond to subject nodes a, b, and e that are not connected to persona C2 by edge 32, or the similarity between the persona features of persona d2 that corresponds to persona node d2.

[0035] The recommendation loss calculation unit 105 calculates, for example, the difference between the first similarity and the second similarity obtained for the subject node C and the persona C2 calculated as shown in Figures 5 and 6 as the loss of the subject node C. The recommendation loss calculation unit 105 may calculate the average value of the losses of all or some of the subject nodes as the recommendation loss.

[0036] 5 and 6 assume the recommendation loss according to equation (1), but other combinations may be used. For example, if the recommendation loss according to equation (1) is used as the first calculation method, the recommendation loss calculation unit 105 calculates, as the second calculation method, for each subject node, the difference between the first similarity between a first subject node selected from category A30 and a subject node or persona node in category B35 connected by edge 32 and the second similarity between the first subject node and a subject node or persona node in category B35 that is not connected by edge 32. The recommendation loss may be calculated by averaging the differences calculated for each subject node. Specifically, using FIG. 3 as an example, the difference between the first similarity between subject node C and subject nodes a, b, c, and d and persona node d1 connected by edge 32 and the second similarity between subject node e and persona node d2 that are not connected by an edge to subject node C may be calculated.

[0037] Alternatively, as a third calculation method, the recommendation loss calculation unit 105 calculates the difference between the first similarity between the subject node or persona node in category B35 connected to the first persona node selected from category A30 by edge 32 and the second similarity between the subject node or persona node in category B35 that is not connected to the first persona node by edge 32. The recommendation loss may also be calculated by averaging the differences calculated for each persona node. Specifically, using FIG. 3 as an example, it is sufficient to calculate the difference between the first similarity obtained between persona node C2 and subject nodes c and d and persona node d1 connected by edge 32 and the second similarity obtained between persona node C2 and subject nodes a, b, and e and persona node d2 that are not connected by edge 32.

[0038] Alternatively, as a fourth calculation method, the recommendation loss calculation unit 105 calculates the difference between the first similarity between a subject node or persona node in category B35 connected to a first persona node selected from category A30 by edge 32, and the second similarity between a subject node or persona node in category B35 that is not connected to the first subject node by edge 32. The recommendation loss calculation unit 105 may calculate the recommendation loss by averaging the differences calculated for each persona node. Specifically, using FIG. 3 as an example, it is sufficient to calculate the difference between the first similarity obtained between persona node C2 and subject nodes c and d and persona node d1 connected by edge 32, and the second similarity obtained between subject node C and subject node e and persona node d2 that are not connected by an edge.

[0039] Furthermore, the recommendation loss calculation unit 105 may calculate the recommendation loss using a combination of at least two of the first to fourth calculation methods. For example, the recommendation loss calculation unit 105 may determine the recommendation loss as a weighted sum of the calculation methods.

[0040] Note that, as a recommendation loss from another perspective, the recommendation loss calculation unit 105 may calculate the difference between the first similarity between the first persona node selected from category A30 and a subject node or persona node in category B35 that is not connected by edge 32, and the second similarity between the first subject node and a subject node or persona node in category B35 that is not connected by edge 32. The recommendation loss calculation unit 105 may calculate an additional recommendation loss (also referred to as a third loss) using the average value of the differences calculated for each first persona node. The additional recommendation loss is a loss for taking into consideration the characteristics of products that neither the user nor the persona has purchased. The update unit 107 updates the parameters of the model based on the recommendation loss, the additional recommendation loss, and the comparison loss.

[0041] Next, the relationship between nodes when calculating the comparison loss based on the additional graph data will be described with reference to FIG. FIG. 7 shows additional graph nodes similar to those in FIGS. 5 and 6, with the subject nodes or persona nodes adjacent to subject node C and persona node C1 shown in bold. The comparison loss calculation unit 106 calculates the difference between the first similarity between the first subject node selected from category A30 and the subject node or persona node of category B35 connected by edge 32, and the second similarity between the first persona node and the subject node or persona node of category B35 connected by edge 32.

[0042] Specifically, the comparison loss calculation unit 106 calculates the similarity between persona node C1 and subject nodes a and b connected by edge 32 as the first similarity, and calculates the similarity between subject node C and subject nodes a, b, c, and d and persona node d1 connected by edge 32 as the second similarity. The comparison loss calculation unit 106 may calculate the difference between the first similarity and the second similarity, and calculate the comparison loss using the average value of the differences calculated for each persona node.

[0043] According to the present embodiment described above, additional graph data is generated by adding personas related to the subject node as nodes, and recommendation loss and comparison loss are calculated from the subject feature and persona feature. Based on the recommendation loss and comparison loss, the parameters of a model that extracts features from the graph data are updated. This allows efficient training of relationships between subjects. For example, when training a model for the relative purchase probability between a user and a product, i.e., for rankings, the model is trained to separate the features between the user and the persona, thereby training the model for the relative relationship between the user and the persona, i.e., which is more likely to purchase. Furthermore, by separating the features between the user and the persona, the interpretability of the model, i.e., which persona is more likely to purchase a product, can be improved. In other words, a trained model with improved interpretability can be provided.

[0044] An example of the hardware configuration of the information processing device 10 according to the above embodiment will now be described with reference to the block diagram of FIG. The information processing device 10 includes a CPU (Central Processing Unit) 81, a RAM (Random Access Memory) 82, a ROM (Read Only Memory) 83, a storage 84, a display device 85, an input device 86, and a communication device 87, each of which is connected by a bus.

[0045] The CPU 81 is a processor that executes arithmetic processing, control processing, etc. in accordance with a program. The CPU 81 uses a predetermined area of ​​the RAM 82 as a working area and executes the processing of each part of the information processing device 10 and the server 11 in cooperation with programs stored in the ROM 83, the storage 84, etc.

[0046] The RAM 82 is a memory such as an SDRAM (Synchronous Dynamic Random Access Memory), and functions as a work area for the CPU 81. The ROM 83 is a memory that stores programs and various information in a non-rewritable manner.

[0047] The storage 84 is a device that writes and reads data to and from a magnetic recording medium such as a hard disk drive (HDD), a semiconductor storage medium such as a flash memory, or an optically recordable storage medium, etc. The storage 84 writes and reads data to and from the storage medium in accordance with control from the CPU 81.

[0048] The display device 85 is a display device such as an LCD (Liquid Crystal Display), etc. The display device 85 displays various information based on a display signal from the CPU 81. The input device 86 is an input device such as a mouse, a keyboard, etc. The input device 86 receives information input by a user as an instruction signal, and outputs the instruction signal to the CPU 81. The communication device 87 communicates with external devices via a network under the control of the CPU 81 .

[0049] The instructions shown in the processing procedures described in the above-described embodiments can be executed based on a software program. A general-purpose computer system can store this program in advance and, by loading this program, achieve effects similar to those achieved by the control operations of the learning system (local device and server) described above. The instructions described in the above-described embodiments can be recorded as a computer-executable program on a magnetic disk (such as a flexible disk or hard disk), an optical disk (such as a CD-ROM, CD-R, CD-RW, DVD-ROM, DVD±R, DVD±RW, or Blu-ray Disc), a semiconductor memory, or a similar recording medium. The recording medium may take any storage format as long as it is readable by a computer or embedded system. A computer can load the program from the recording medium and execute the instructions described in the program on a CPU based on the program, thereby achieving operations similar to those of the learning system (local device and server) described in the above-described embodiments. Of course, the computer may acquire or load the program via a network. In addition, an OS (operating system), database management software, network middleware, etc. running on a computer may execute some of the processes required to realize this embodiment based on instructions from a program installed on the computer or embedded system from a recording medium.

[0050] Furthermore, the recording medium in this embodiment is not limited to a medium independent of a computer or an embedded system, but also includes a recording medium that stores or temporarily stores a program downloaded via a LAN, the Internet, or the like. Furthermore, the number of recording media is not limited to one, and cases where the processing in this embodiment is executed from multiple media are also included in the recording media in this embodiment, and the media may have any configuration.

[0051] The computer or embedded system in this embodiment is for executing each process in this embodiment based on a program stored on a recording medium, and may be configured as either a device consisting of a single device such as a personal computer or a microcomputer, or a system in which multiple devices are connected to a network. Furthermore, the computer in this embodiment is not limited to a personal computer, but also includes an arithmetic processing unit, a microcomputer, etc. included in information processing equipment, and is a general term for equipment or devices that can realize the functions in this embodiment by a program.

[0052] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0053] 10 Information processing device, 30 Subject A category, 31 Subject node, 32 Edge, 33 Persona node, 35 Subject B category, 40, 45 Community, 81 CPU, 82 RAM, 83 ROM, 84 Storage, 85 Display device, 86 Input device, 87 Communication device, 101 Storage unit, 102 Acquisition unit, 103 Persona generation unit, 104 Extraction unit, 105 Recommendation loss calculation unit, 106 Comparison loss calculation unit, 107 Update unit.

Claims

1. an acquisition unit that acquires graph data in which a plurality of entities are represented by nodes for each of two or more categories and relationships between the entities are represented by edges connecting the nodes; a persona generation unit that generates additional graph data to which one or more personas having representative characteristics of the subject are added as the nodes; an extraction unit that extracts subject features and persona features from the additional graph data using a model that extracts features; a recommendation loss calculation unit that calculates a first loss according to a difference between a similarity between nodes connected by the edge and a similarity between nodes not connected by the edge, using the subject feature and the persona feature; a comparison loss calculation unit that calculates a second loss according to a difference between a similarity between the subject node and a node connected to the subject node and a similarity between the persona node and a node connected to the persona node, using the subject feature amount and the persona feature amount; an update unit that updates the model using the first loss and the second loss; An information processing device comprising:

2. The information processing device according to claim 1 , wherein the update unit updates the parameters of the model so that a difference in similarity between the first loss and the second loss increases.

3. The persona is generated for each of the entities; 2. The information processing device according to claim 1, wherein the recommendation loss calculation unit calculates the first loss according to a difference between a similarity between a node of a first subject among the plurality of subjects and a node of another subject or the persona connected by the edge, and a similarity between a node of a first persona among the plurality of personas and a node of the subject or another persona not connected by the edge.

4. 2. The information processing device according to claim 1, wherein the recommendation loss calculation unit calculates the first loss according to a difference between a similarity between a node of a first subject among the plurality of subjects and a node of another subject or the persona connected by the edge, and a similarity between the node of the first subject and a node of the other subject or the persona not connected by the edge.

5. The persona is generated for each of the entities; 2. The information processing device according to claim 1, wherein the recommendation loss calculation unit calculates the first loss according to a difference between a similarity between a node of a first persona among a plurality of personas and a node of the subject or another persona connected by the edge, and a similarity between a node of the first persona and a node of the subject or another persona that is not connected by the edge.

6. The persona is generated for each of the entities; 2. The information processing device according to claim 1, wherein the recommendation loss calculation unit calculates the first loss according to a difference between a similarity between a node of a first persona among a plurality of personas and a node of the subject or another persona connected by the edge, and a similarity between a node of the first subject among the plurality of subjects and a node of another subject or the persona that is not connected by the edge.

7. 2. The information processing device according to claim 1, wherein the recommendation loss calculation unit calculates the first loss according to at least two combinations of the difference between the similarity between the entities or the personas connected by the edge and the similarity between the entities or the personas not connected by the edge.

8. The persona is generated for each of the entities; the recommendation loss calculation unit calculates a third loss according to a difference between a similarity between a node of a first persona among a plurality of personas and a node of the subject or another persona that is not connected by the edge, and a similarity between a node of the first subject among the plurality of subjects and a node of the other subject or the persona that is not connected by the edge; The information processing device according to claim 1 , wherein the update unit updates the model using the first loss, the second loss, and the third loss.

9. The information processing device according to claim 1 , wherein the graph data represents a purchasing relationship between a person and a product, or a visiting relationship between a person and a place.

10. The information processing device according to claim 1 , wherein the extraction unit extracts the subject features and the persona features using a graph neural network having a message passing mechanism, such as a graph convolutional network, GAT, GraphSAGE, or GIN.

11. The information processing device according to claim 1 , wherein the persona generation unit generates the persona using an egocentric network, a graph Laplacian, and Ricci curvature.

12. The acquiring means acquires graph data in which a plurality of entities are represented by nodes for each of two or more categories, and relationships between the entities are represented by edges connecting the nodes; a persona generation means for generating additional graph data in which one or more personas having representative characteristics of the subject are added as the nodes; an extraction means extracting subject features and persona features from the additional graph data using a model for extracting features; a recommendation loss calculation means, using the subject feature and the persona feature, calculates a first loss according to a difference between a similarity between the nodes connected by the edge and a similarity between the nodes not connected by the edge; a comparison loss calculation means, using the subject feature amount and the persona feature amount, calculates a second loss according to a difference between a similarity between the subject node and a node connected to the subject node and a similarity between the persona node and a node connected to the persona node; An information processing method, wherein an update means updates the model using the first loss and the second loss.

13. Computer, an acquisition means for acquiring graph data in which a plurality of entities are represented by nodes for each of two or more categories and relationships between the entities are represented by edges connecting the nodes; a persona generation means for generating additional graph data in which one or more personas having representative characteristics of the subject are added as the nodes; an extraction means for extracting subject features and persona features from the additional graph data using a model for extracting features; a recommendation loss calculation means for calculating a first loss according to a difference between a similarity between nodes connected by the edge and a similarity between nodes not connected by the edge, using the subject feature and the persona feature; a comparison loss calculation means for calculating a second loss according to a difference between a similarity between the subject node and a node connected to the subject node and a similarity between the persona node and a node connected to the persona node, using the subject feature amount and the persona feature amount; an information processing program for causing the model to function as an update unit that updates the model using the first loss and the second loss;

Images