Display method, recommendation method and recommendation device

Abstract

The invention discloses a display method, a recommendation method and a recommendation device, and the method comprises the steps: determining a relation intention and a context intention of a user based on historical interaction data of the user; the relationship intention represents the degree of attention of the user to the relationship between the interacted article and the associated entity, and the associated entity comprises the entity having the relationship in the knowledge graph of the article; the context intention represents the attention degree of the user to the interaction context; based on the relationship intention and the context intention of the user, obtaining embedded representation information through a graph neural network; and the recommendation model outputs the second display content at least according to the embedded representation information. In the application, the intention of the user is mined through the historical interaction data of the user to obtain the relationship intention and the context intention. The embedded representation information of the user comprises a relation intention and a context intention of the user, so that the embedded representation information is an accurate representation. Therefore, when the recommendation model takes the embedded representation information as input data, a more accurate recommendation result can be obtained.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a display method, a recommendation method and a recommendation device. Background Art

[0002] In the field of artificial intelligence technology, recommendation services are an important tool to reduce information overload and improve user experience. As the amount of data continues to increase, recommendation systems are playing an increasingly important role. An important issue that recommendation systems need to solve is how to improve the recommendation effect and provide users with more accurate information from massive data. Summary of the invention

[0003] In order to solve the above-mentioned problems, a display method, a recommendation method and a recommendation device are provided in the embodiments of the present application to determine the contextual intent related to the context of the user's interaction with an item and the relational intent related to the relationship between the interacted items and the associated entities in the knowledge graph, and obtain embedded representation information representing the contextual intent and the relational intent through a graph neural network, and embed the user's embedded representation information into the input layer of the recommendation model. The recommendation model can obtain more accurate recommendation results under the input data with refined representation, thereby improving the recommendation effect of the recommendation system, so that the recommendation system can output more accurate prediction information.

[0004] To this end, the following technical solutions are adopted in the embodiments of the present application:

[0005] In the first aspect, an embodiment of the present application provides a display method, including: a first device displays a first user interface, the first user interface includes: a first display content; after the first device detects a first operation acting on the first display content, it displays a second user interface; the second user interface includes: a second display content; the second display content is related to the relational intent and the contextual intent; wherein the second display content is the result of the recommendation model making a recommendation based at least on embedded representation information; the embedded representation information is the representation information output by the graph neural network based on the relational intent and the contextual intent; the relational intent represents the user's attention to the relationship between the first display content and the associated entities, and the associated entities include entities with relationships in the knowledge graph of the first display content; the contextual intent represents the user's attention to the context of the first operation.

[0006] In an embodiment of the present application, based on the user's first operation on one or more first displayed contents, the user's intention is mined to obtain relational intention and contextual intention. Among them, the relationship between the item and the associated entity is determined based on the knowledge graph of the item interacted by the user; based on the user's attention to the relationship, the user's relational intention is further determined. Contextual intention represents the user's attention to the context of the interaction. After passing through the graph neural network, the user's embedded representation information is obtained. The embedded representation information contains the user's relational intention and contextual intention. Therefore, the embedded representation information is an accurate representation of the user. The recommendation model can output accurate recommendation results based on the user's embedded representation information. In other words, the embodiment of the present application inputs accurate data to the recommendation model so that the recommendation model can output accurate recommendation results. In the training stage, accurate input data and accurate recommendation results can be trained to obtain an accurate recommendation model or recommendation system. In the prediction stage, accurate input representation data can be used to accurately represent the predicted sample, thereby outputting accurate recommendation results and improving user experience.

[0007] In one possible implementation, after the first device displays the second user interface, the display method also includes: after the first device detects the second operation, displaying a third user interface; the third user interface displays: at least one associated entity, the user's attention level to the associated entity; and / or at least one context of the first operation, the user's attention level to the context of the first operation.

[0008] In this implementation, after the user performs the second operation, the first device displays the third user interface, and the third user interface displays at least one of the relationship intent and the context intent. The display of the relationship intent may be at least one associated entity and the user's attention to the associated entity. The context intent may be the context of at least one first operation and the user's attention to the context of the first operation. The user can understand the reason for recommending the second displayed content through the third user interface, which improves the user experience and makes the recommendation results more explainable.

[0009] In another possible implementation, after the first device displays the second user interface, the display method also includes: after the first device detects a third operation, displaying a fourth user interface; the fourth user interface displays: at least one first display content, and / or at least one associated entity, and / or at least one context of the first operation.

[0010] In this implementation, after the user performs the third operation, the first device displays the fourth user interface, and the fourth user interface displays the user's historical interaction data, that is, the relevant information of the first operation and the first display content. Among them, the relevant information of the first operation may be the context of the first operation. The relevant information of the first display content may be at least one first display content and / or at least one associated entity. Through the fourth user interface, the user can understand the historical interaction data, the items of historical interaction, the associated entities, the context of the interaction, and other information, which facilitates the user's recording and query of historical interaction data, and improves the user experience.

[0011] In another possible implementation, the embedded representation information is the representation information output by the graph neural network based on relational intent and contextual intent, including: the graph representation input to the graph neural network includes user nodes, item nodes, fusion intent nodes, and connecting edges; the user node represents the user, the item node represents the first display content, and the fusion intent node represents the user's fusion intent; the fusion intent represents the result of the fusion of relational intent and contextual intent; the connecting edges are used to connect the user node and the fusion intent node, and the fusion intent node and the item node; the graph neural network is used to update the features of the nodes represented by the graph and output the embedded representation information; the embedded representation information includes the user representation output by the user node and the item representation output by the item node.

[0012] In this implementation, the graph representation of the input graph neural network includes user nodes, item nodes, fusion intent nodes, and connection edges. The graph neural network propagates information on the nodes of the graph representation, and the output embedded representation information of the user includes the user representation output by the user node and the item representation output by the item node. In other words, the user's precise embedded representation information includes precise user representation and precise item representation. Therefore, when the input of the recommendation model is precise user representation and precise item representation, the recommendation model can output precise recommendation results. Furthermore, by embedding contextual intent and relational intent into the fusion intent node, when introducing the knowledge graph of the item, it can avoid adding a large number of heterogeneous nodes, reduce the complexity of the calculation, and avoid introducing too much noise.

[0013] In another possible implementation, features of the nodes represented by the graph are updated, including: based on features of the item node and the fusion intent node, features of the user node are updated; based on features of the user node and the fusion intent node, features of the item node are updated.

[0014] In this implementation, the fusion intention node is used as the hub to update the features of the user's relational intention and contextual intention to the user node and item node. Therefore, the representation of the user node and item node output by the graph neural network contains the information of the relational intention and contextual intention. In other words, after the graph representation is updated by the graph neural network, the user node outputs an accurate user representation and the item node outputs an accurate item representation. By setting a fusion intention node between the user node and the item node, the propagation distance of the relational intention to the user node is reduced, ensuring the influence of the relational intention on the user's embedded representation information.

[0015] In another possible implementation, the graph representation of the graph neural network input also includes context intent nodes and context nodes; the connecting edges are also used to connect user nodes and context intent nodes, context intent nodes and context nodes, and item nodes and context intent nodes; the context intent nodes represent the context intent, and the context nodes represent the context of the first operation; the embedded representation information also includes the context representation of the context node output.

[0016] In this implementation, another graph representation of the input graph neural network is provided. The graph representation includes user nodes, item nodes, fusion intent nodes, context intent nodes, context nodes, and connection edges. That is, the nodes representing user intent are refined into fusion intent nodes and context intent nodes. Fusion intent nodes are used to embed fusion intents fused with relationship intents and context intents. Context intent nodes are used to embed context intents. The graph representation sets context nodes, and by continuously updating the features of the context nodes, an accurately characterized context representation can be obtained. In other words, this implementation can obtain accurate user representations, item representations, and context representations, further improving the input data of the recommendation model, so that the recommendation model can output more accurate recommendation results.

[0017] In another possible implementation, information propagation is performed on the nodes represented by the graph, including: updating the features of the user node based on the features of the item node and the fused intent node; updating the features of the item node based on the features of the user node and the fused intent node; after updating the features of the context intent node based on the information of the user node and the item node, updating the features of the context node based on the context intent and the features of the context intent node; and / or updating the features of the context intent node based on the features after the features of the context node are updated, so that the features of the fused intent node are updated based on the features of the context intent node.

[0018] In this implementation, by updating the features of the user node and the item node, the user representation output by the user node and the item representation output by the item node can contain the user's fusion intention, thereby making the user representation output by the user node and the item representation output by the item node more refined. By updating the features of the context node, the embedded representation of the context node can contain the features of the item node and the user node, making the context representation output by the context node more refined. The refined context representation can be further propagated to the context intent node, and then through the context intent node, the refined context representation is used to update the features of other nodes.

[0019] In another possible implementation, the graph representation also includes entity nodes and relationship edges; the relationship edges are used to connect item nodes and entity nodes; the entity nodes represent associated entities; the relationship edges represent the relationship between items and associated entities; the initial features of the entity nodes are the information of the associated entities in the knowledge graph; updating the features of the nodes represented by the graph also includes: updating the features of the entity nodes based on the features of the item nodes and the relationship edges; updating the features of the item nodes based on the features of the entity nodes and the relationship edges.

[0020] In this implementation, an information propagation process including entity nodes and relationship connection edge graph representation is provided. That is, entity nodes can also receive information from adjacent nodes, making the embedding representation of entity nodes more accurate. The refined embedding representation of entity nodes can be propagated to relationship intent and fusion intent, so that when relationship intent and fusion intent are propagated to user nodes, item nodes, and context nodes, more accurate embedding representations of users, items, and contexts can be obtained.

[0021] In another possible implementation, the relationship intent represents the user's attention to the relationship between the first displayed content and the associated entity, including: obtaining the relationship intent based on the user's attention to one or more relationship sub-intentions; the relationship sub-intention represents the attention to the associated entity.

[0022] In this implementation, a method for calculating relational intent is provided. Through the knowledge graph, items, associated entities, and the relationship between items and associated entities are obtained, and the relational sub-intent is obtained based on the degree of association of the relationship. In other words, the relational sub-intent represents the degree of association of the associated entity. The relational sub-intent can be understood as a rough classification of the user's relational intent. For example, suppose the item is restaurant 1, the associated entity is location 1, front desk service 2, and price 3. In the first relational sub-intent, the degrees of association of the three associated entities are 0.5, 0.3, and 0.2, respectively. In the second relational sub-intent, the degrees of association of the three entities are 0.2, 0.2, and 0.6, respectively. If the user is more concerned about the price, the weight of the first relational sub-intent in the user's relational intent is greater. If the user is more concerned about the location, the weight of the first relational sub-intent in the user's relational intent is greater. If the user is more concerned about the front desk service, the weights of the first relational sub-intent and the second relational sub-intent in the user's relational intent are roughly equal. In actual application scenarios, the types and quantities of items and associated entities are huge and complex. Therefore, the relationship sub-intent can be determined based on the items and associated entities. Adding the calculation of attention on the basis of the relationship sub-intent to determine the relationship intent can greatly reduce the amount of calculation and improve the calculation efficiency of the relationship intent. In other words, the relationship intent is the degree of attention of the user to the relationship sub-intent, and is the personalized embodiment of the user on the relationship sub-intent. Therefore, after the relationship intent is propagated to the user node, the user representation, item representation, and context representation obtained improve the distinction between the user's interaction data and make the granularity of the embedded representation higher. In other words, the relationship sub-intent is a shared intent. The relationship sub-intent represents the degree of association between the relationship between the item and the entity, which is irrelevant to the user and does not change due to the change of the user. Exemplarily, based on the similarity between the user and the relationship sub-intent, the weight of the relationship sub-intent in the user's relationship intent is determined, that is, the degree of attention of the user to the relationship sub-intent, and then the relationship intent is obtained. Therefore, the relationship sub-intent is a shared intent, which reduces the amount of calculation and reduces the complexity of the model.

[0023] In another possible implementation, the contextual intent further represents the user's attention to the context of the fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

[0024] In this implementation, the context of the first operation includes not only the context of the current interaction event, but also the context of the previous interaction event. Therefore, the embedded representation of user, item, and context information includes information about the current interaction event and historical interaction events, and adds information about the context of the user's historical interactions. In other words, since the user's preferences will not change in a short period of time, the embedded representation of user, item, and context is related to the context information of historical interactions, which can reduce the error of the recommendation results and further improve the refinement of user representation, item representation, and context representation.

[0025] In the second aspect, an embodiment of the present application provides a recommendation method, including: determining the user's relational intention and contextual intention based on the user's historical interaction data; the historical interaction data includes the user's first operation on the first display content and the context of the first operation; the relational intention represents the user's attention to the relationship between the first display content and the associated entities, and the associated entities include entities with relationships in the knowledge graph of the first display content; the contextual intention represents the user's attention to the context of the first operation; based on the user's relational intention and contextual intention, embedded representation information is obtained through a graph neural network; the recommendation model outputs the second display content based on at least the embedded representation information.

[0026] In a possible implementation, based on the user's relational intention and contextual intention, embedded representation information is obtained through a graph neural network, including: inputting the graph representation into the graph neural network; the graph representation includes user nodes, item nodes, fusion intention nodes, and connecting edges; the user node represents the user, the item node represents the first display content, and the fusion intention node represents the user's fusion intention; the fusion intention represents the result of the fusion of relational intention and contextual intention; the connecting edge is used to connect the user node and the fusion intention node, and the fusion intention node and the item node; the graph neural network updates the features of the nodes of the graph representation and outputs the embedded representation information; the embedded representation information includes the user representation output by the user node and the item representation output by the item node.

[0027] In another possible implementation, the graph neural network updates the features of the nodes represented by the graph, including: updating the features of the user nodes based on the features of the item nodes and the fusion intent nodes; and updating the features of the item nodes based on the features of the user nodes and the fusion intent nodes.

[0028] In another possible implementation, based on the user's relational intention and contextual intention, embedded representation information is obtained through a graph neural network, including: the graph representation also includes contextual intention nodes and context nodes; the connecting edges are also used to connect user nodes and contextual intention nodes, contextual intention nodes and context nodes, and item nodes and contextual intention nodes; the contextual intention nodes represent the user's contextual intention, and the context nodes represent the context of the first operation; the embedded representation information also includes the context representation of the context node output.

[0029] In another possible implementation, the graph neural network updates features of the nodes represented by the graph, including: updating features of user nodes based on features of item nodes and fused intent nodes; updating features of item nodes based on features of user nodes and fused intent nodes; updating features of context intent nodes based on features of user nodes and item nodes, and then updating features of context nodes based on context intent and features of context intent nodes; and / or updating features of context intent nodes based on features after feature updates of context nodes, so that features of fused intent nodes are updated based on features of context intent nodes.

[0030] In another possible implementation, the graph representation also includes entity nodes and relationship edges; entity nodes represent associated entities; relationship edges represent the relationship between items and associated entities; the initial features of entity nodes are the information of associated entities in the knowledge graph; relationship edges are used to connect item nodes and entity nodes; the graph neural network updates the features of the nodes represented by the graph, and also includes: based on the features of item nodes and relationship edges, feature updates of entity nodes; based on the features of entity nodes and relationship edges, feature updates of item nodes.

[0031] In another possible implementation, relationship intent represents the user's attention to the relationship between an item and an associated entity, including: obtaining relationship intent based on the user's attention to one or more relationship sub-intentions; relationship sub-intention represents the attention to the associated entity.

[0032] In another possible implementation, the contextual intent further represents the user's attention to the context of the fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

[0033] In a third aspect, an embodiment of the present application provides a recommendation device, comprising: an intention determination module, for determining a user's relational intention and contextual intention based on the user's historical interaction data; the historical interaction data includes the user's first operation on a first displayed content and the context of the first operation; the relational intention represents the user's attention to the relationship between the first displayed content and the associated entities, and the associated entities include entities with relationships in the knowledge graph of the first displayed content; the contextual intention represents the user's attention to the context of the first operation; the representation determination module, for obtaining embedded representation information based on the user's relational intention and contextual intention through a graph neural network; and the recommendation module, for outputting a second display content through a recommendation model at least based on the embedded representation information.

[0034] In one possible implementation, the representation determination module is specifically used to: input the graph representation into a graph neural network; the graph representation includes user nodes, item nodes, fusion intent nodes, and connection edges; the user node represents the user, the item node represents the first display content, and the fusion intent node represents the user's fusion intent; the fusion intent represents the result of the fusion of relationship intent and context intent; the connection edge is used to connect the user node and the fusion intent node, the fusion intent node and the item node; the graph neural network is used to update the features of the nodes represented by the graph, and output embedded representation information; the embedded representation information includes the user representation output by the user node and the item representation output by the item node.

[0035] In another possible implementation, the characterization determination module is specifically used to: update the features of the user node based on the features of the item node and the fusion intention node; and update the features of the item node based on the features of the user node and the fusion intention node.

[0036] In another possible implementation, the representation determination module is specifically used to: the graph representation also includes context intent nodes and context nodes; the connecting edges are also used to connect user nodes and context intent nodes, context intent nodes and context nodes, item nodes and context intent nodes; the context intent nodes represent the user's context intent, and the context nodes represent the context of the first operation; the embedded representation information also includes the context representation of the context node output.

[0037] In another possible implementation, the characterization determination module is specifically used to: update the features of the user node based on the features of the item node and the fused intent node; update the features of the item node based on the features of the user node and the fused intent node; after updating the features of the context intent node based on the features of the user node and the item node, update the features of the context node based on the context intent and the features of the context intent node; and / or update the features of the context intent node based on the features after the features of the context node are updated, so that the features of the fused intent node are updated based on the features of the context intent node.

[0038] In another possible implementation, the graph representation also includes entity nodes and relationship edges; entity nodes represent associated entities; relationship edges represent the relationship between items and associated entities; the initial features of the entity nodes are the information of the associated entities in the knowledge graph; relationship edges are used to connect item nodes and entity nodes; the representation determination module is specifically used to: update the features of the entity nodes based on the features of the item nodes and relationship edges; update the features of the item nodes based on the features of the entity nodes and relationship edges.

[0039] In another possible implementation, the intent determination module is specifically used to: obtain a relationship intent based on the user's attention to one or more relationship sub-intents; the relationship sub-intent represents the degree of attention to the associated entity.

[0040] In another possible implementation, the contextual intent further represents the user's attention to the context of the fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

[0041] In a fourth aspect, the present application provides a computing device, including a memory and a processor. The memory stores instructions, and when the instructions are executed by the processor, the above-mentioned display method or recommendation method or recommendation device is implemented.

[0042] In a fifth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the above-mentioned display method or recommendation method or recommendation device is implemented.

[0043] In a sixth aspect, the present application provides a computer program product, which includes program instructions. When the program instructions are executed by a computer, the computer executes the above-mentioned display method or recommendation method, or realizes the functions of the above-mentioned recommendation device. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The following is a brief introduction to the drawings required for use in the embodiments or technical descriptions.

[0045] Figure 1 A schematic diagram of an application scenario of a recommended method provided in an embodiment of the present application;

[0046] Figure 2 A schematic diagram of a display method provided in an embodiment of the present application;

[0047] Figure 3 A schematic diagram of a user interface 10 provided in an embodiment of the present application;

[0048] Figure 4 A schematic diagram of a user interface 20 provided in an embodiment of the present application;

[0049] Figure 5a A schematic diagram of a user interface 30 provided in an embodiment of the present application;

[0050] Figure 5b A schematic diagram of another user interface 30 provided in an embodiment of the present application;

[0051] Figure 6 A schematic diagram of a user interface 40 provided in an embodiment of the present application;

[0052] Figure 7 A schematic diagram of a recommended method provided in an embodiment of the present application;

[0053] Figure 8 A schematic diagram of a combination of a personal graph and a knowledge graph provided in an embodiment of the present application;

[0054] Fig. 9 A schematic diagram of an implementation method of the relationship intention provided in the embodiments of the present application;

[0055] Fig.10 A schematic diagram of an implementation method of a context graph provided in an embodiment of the present application;

[0056] Fig.11 A schematic diagram of an implementation method of the fusion intention provided in an embodiment of the present application;

[0057] Fig.12 A schematic diagram of a PING graph structure including a relationship intent node and a context intent node, or including a fusion intent node, provided in an embodiment of the present application;

[0058] Fig.13 A schematic diagram of a PING graph structure including context nodes and context intent nodes provided in an embodiment of the present application;

[0059] Fig.14A schematic diagram of a PING graph structure including entity nodes and relationship connection edges provided in an embodiment of the present application;

[0060] Fig.15 A schematic diagram of a PING graph structure including multiple users provided in an embodiment of the present application;

[0061] Fig.16 A schematic diagram of a graph neural network providing in an embodiment of the present application performing feature update on a PING graph structure;

[0062] Fig.17 A schematic diagram of information propagation to a user node in a graph neural network provided in an embodiment of the present application;

[0063] Fig.18 A schematic diagram of information propagation to an item node in a graph neural network provided in an embodiment of the present application;

[0064] Fig.19 A schematic diagram of information propagation to a context node in a graph neural network provided in an embodiment of the present application;

[0065] Fig. 20 A schematic diagram of information propagation to entity nodes in a graph neural network provided in an embodiment of the present application;

[0066] Fig.21 A schematic diagram of a recommendation model provided in an embodiment of the present application;

[0067] Fig. 22 A schematic diagram of the composition of a recommendation system architecture provided in an embodiment of the present application;

[0068] Fig.23 A schematic diagram of a node accessing high-order neighbor information provided in an embodiment of the present application;

[0069] Fig.24 A schematic diagram of the composition of a recommended device provided in an embodiment of the present application;

[0070] Fig.25 A schematic diagram of the composition of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0072] The term "and / or" in this article is a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The symbol " / " in this article indicates that the associated objects are in an "or" relationship, for example, A / B means A or B.

[0073] The terms "first" and "second" in the specification and claims herein are used to distinguish different objects rather than to describe a specific order of the objects. For example, a first response message and a second response message are used to distinguish different response messages rather than to describe a specific order of the response messages.

[0074] In the embodiments of the present application, the words "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0075] In the description of the embodiments of the present application, unless otherwise specified, "multiple" means two or more than two. For example, multiple processing units refer to two or more processing units, etc.; multiple elements refer to two or more elements, etc.

[0076] In order to facilitate understanding of the solution provided by the embodiment of the present application, some terms involved in the solution are first briefly introduced.

[0077] Knowledge Graph (KG): Stores structured information of real-world facts, such as project attributes, taxonomies, or external common sense knowledge, in the form of heterogeneous graphs or heterogeneous information networks. A knowledge graph is a heterogeneous graph or heterogeneous information network, including nodes and connecting edges. In a knowledge graph, nodes represent entities, represented by entity sets V, and connecting edges represent the relationships between entities, represented by relationship sets R. Exemplarily, in a knowledge graph, let (h, r, t) be the triple of the knowledge graph, h (entity head) and t (entity tail) belong to the entity set V, and r belongs to the relationship set R. For example, (Hotel A, provides, outdoor seating) describes that Hotel A provides outdoor seating.

[0078] Item: refers to an object or entity, such as items that users interact with, items recommended to users, etc.

[0079] Related entities: In the knowledge graph, entities that have a relationship with items are called related entities. For example, if the item that the user interacts with is Hotel A, in the knowledge graph of Hotel A, the related entities of Hotel A include: outdoor chairs, hotel front desk service, music, etc.

[0080] Relationship intent: represents the degree of attention users pay to the relationship between items and related entities. For example, the relationship between hotel A and outdoor chairs is provision, the relationship between hotel A and hotel front desk service is provision and good quality, and the relationship between hotel A and music is free provision. Therefore, when users are more concerned about provision of outdoor chairs, the relationship intent between hotel A and outdoor chairs will be greater.

[0081] Relationship sub-intention: A shared intention that can calculate the user's relationship intention based on the user's attention to the relationship sub-intention, avoiding the calculation of the user's relationship intention based on the relationship between the item and the associated entity, reducing the amount of calculation. Figure 1 Represents the intention of paying attention to the stay experience. You can set a higher degree of attention to the relationship between related entities such as outdoor seats, hotel front desk services, and music. Figure 2 To represent the intention of paying attention to price, a higher degree of attention can be set for the relationship of related entities such as hotel price and free breakfast. When calculating the relationship intention of user 1, the relationship sub-intention Figure 1 and relationship Figure 2 The attention mechanism is performed to obtain the relationship intention of user 1. Therefore, when calculating the relationship intention of user 1, it is not necessary to perform attention calculation on the relationship of each associated entity, which reduces the amount of calculation and improves the calculation efficiency of the recommendation system.

[0082] Context: refers to relevant information that affects user interaction. For example, environmental information, such as the temperature and climate when the user selects Hotel A; time information, such as the interaction time when the user selects Hotel A; location information, such as the interaction location when the user selects Hotel A. Among them, the environmental information of the interaction can be high temperature, rainfall, etc.; the interaction time can be the year, season, week, etc.; the interaction location can be the city, etc. In other words, in the embodiment of the present application, context refers to information that helps or affects the user to make a decision.

[0083] Contextual intent: refers to the degree of attention paid to the contextual information of user interaction. For example, the user selected Hotel A on Wednesday and Hotel B on the weekend. In the user's association matrix for Hotel A, Wednesday will have a larger weight. When the user selects a hotel again, if the current time is Wednesday, the user will be recommended a hotel similar to Hotel A; if the current time is the weekend, the user will be recommended a hotel similar to Hotel B.

[0084] Fusion intent: refers to the intent that is the result of the fusion of relational intent and contextual intent. For example, user 1 chooses restaurant 1 in the morning, and restaurant 1 has free drinks. User 1 chooses restaurant 2 at noon, and restaurant 2 provides music. Among them, the contexts of morning and noon represent the user's contextual intent; free drinks and providing music represent the user's relational intent. Therefore, the fusion intent of user 1 is "in the morning, user 1 tends to choose restaurants with free drinks" and "at noon, user 1 tends to choose restaurants that provide music." When the user makes a restaurant selection in the morning, restaurants with free drinks can be recommended to the user.

[0085] Graph representation: It is a graph structure used for graph neural network input. Graph representation consists of nodes and edges. An edge connects two different nodes. Therefore, multiple edges and multiple nodes form a topological graph structure. For example, if user 1 interacts with item 1, there is an edge between the node representing user 1 and the node representing item 1. In other words, the information of a node in the graph representation represents the connection relationship between the node and the adjacent nodes. One way to represent it is the association matrix.

[0086] Node: In a graph structure, a node is used to represent information about something. For example, a user node represents information about a user. Information about a user includes the relationship between the user and adjacent nodes.

[0087] Edges: In a graph structure, edges are used to connect a node to its neighboring nodes. Edges can be undirected edges. For example, if there is an undirected edge between the node of user 1 and the node of item 1, it means that the information of the node of user 1 and the node of item 1 can be transmitted to each other. Edges can be directed edges. For example, if there is a unidirectional edge between paper 1 and paper 2 pointing to paper 2, it means that paper 2 cites paper 1, but paper 1 does not cite paper 2.

[0088] Graph Neural Networks (GNN): refers to the general term for algorithms that use neural networks to learn graph structure data, extract and discover features and patterns in graph structure data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. In the embodiment of the present application, the graph neural network can update the features of the nodes in the graph representation, so that each node continuously aggregates the information of its neighbors to obtain the representation information of each node.

[0089] Embedded representation information: After the multiple network layers of the graph neural network update the features of the nodes in the graph representation, the representation information of each node is output. Since the representation information of the node aggregates the information of multiple-order neighbors, the content of the representation information of the node is richer and can more accurately represent the things corresponding to the node. In an embodiment of the present application, the representation information of multiple nodes is selected from the representation information of each node output by the graph neural network and embedded in the input layer or embedding layer of the downstream recommendation model, so the representation information of the multiple nodes is called embedded representation information.

[0090] Recommendation system: refers to a system that completes the recommendation task, including a network module that calculates the embedded representation information and a downstream recommendation model. The network module that calculates the embedded representation information calculates and outputs the embedded representation information of multiple users to the downstream recommendation model. Based on the embedded representation information of multiple users, the downstream recommendation model recommends to multiple users. For example, the recommendation model can be matrix factorization (MF), factorization machine (FM), neural network (NN), etc.

[0091] The recommendation method provided in the embodiment of the present application can be applied to scenarios where items need to be recommended, such as hotel recommendation, movie recommendation, product recommendation, music recommendation, restaurant recommendation, book recommendation, web page recommendation, etc. The hotel recommendation scenario is described exemplarily below.

[0092] Figure 1 Schematic diagram of an application scenario of a recommended method provided in an embodiment of the present application. Figure 1 As shown, in an application scenario, user 1 interacts with the display content of device 2. That is, device 2 displays a user interface to user 1, and user 1 operates the display content of the user interface; based on the user's operation, device 2 recommends new display content to user 1. The new display content is related to the context of the user's operation and the associated entities in the knowledge graph of the display content. That is, based on the interaction data between user 1 and the display content, device 2 uses a recommendation method to predict the user's preference and displays the recommended results to the user.

[0093] Optional, such as Figure 1 As shown, the application scenario may also include a server 3. The server predicts the user's preference based on the interaction data between the user and the displayed content using a recommendation method, and displays the recommended result to the user through the device 1.

[0094] Device 1 may be a mobile phone, a tablet computer, a desktop computer, a laptop computer, a handheld computer, a notebook computer, an ultra-mobile personal computer (UMPC), a netbook, a cellular phone, a personal digital assistant (PDA), an augmented reality (AR) device, a virtual reality (VR) device, an artificial intelligence (AI) device, a wearable device, an in-vehicle device, a smart home device and / or a smart city device. The embodiments of the present application do not impose any special restrictions on the specific type of the device.

[0095] It is understood that the structure illustrated in the embodiment of the present invention does not constitute a specific limitation on the device 1. In other embodiments of the present application, the device 1 may include more or fewer components than shown in the figure, or combine some components, or split some components, or arrange the components differently. The components shown in the figure may be implemented in hardware, software, or a combination of software and hardware.

[0096] In an embodiment of the present application, the user interface is used to display images, videos, etc. The user interface may include a variety of software and hardware interfaces, such as a website interface, a graphical user interface, and the like. The user interface may be a liquid crystal display (LCD). The user interface may also be made of an organic light-emitting diode (OLED), an active-matrix organic light-emitting diode or an active-matrix organic light-emitting diode (AMOLED), a flexible light-emitting diode (FLED), miniled, microled, micro-oled, a quantum dot light-emitting diode (QLED), and the like. In some embodiments, the device 1 may include 1 or N user interfaces, where N is a positive integer greater than 1.

[0097] In the embodiment of the present application, the displayed content may be any content, such as text, numbers, symbols, icons, etc., or any combination thereof, such as images, videos, names of items, etc.

[0098] First, in a technical solution of a recommendation method, the knowledge graph is applied to the recommendation system, such as the Knowledge Graph Attention Network (KGAT) and the Knowledge Graph-based Intent Network (KGIN). Further, KGIN is used as an example for explanation. In KGIN, based on the knowledge graph, the different relationships between items and item attributes are used to obtain user intent, and a "user-item-intent" tripartite graph is constructed. Then, the item rating is obtained through a graph convolutional network to learn user interests. In the knowledge graph, item attributes include entities associated with the item and the relationship between the item and the associated entity.

[0099] In this technical solution, the recommendation system based on the knowledge graph ignores the contextual information of user interactions and lacks modeling of the contextual information of user interactions. In other words, in this technical solution, the influence of the user's current and historical contextual information is not considered in the process of predicting user preferences. For example, the KGIN method incorporates the knowledge graph into the model, but does not consider and analyze the contextual features.

[0100] Secondly, the key data source for building recommendation services is the interaction between users and items (e.g., clicks, plays), and interaction data can be used to predict user preferences. User behavior is often affected by contextual information, so a context-aware recommendation system (CARS) is used to integrate contextual information in order to make more accurate and fine-grained recommendations to users.

[0101] In this technical solution, the CARS-related context-aware recommendation method only learns context information, ignoring the relationship between the items interacted by the user and the associated entities of the items.

[0102] Based on the related technical solutions of the first and second aspects, in many cases, the current context and the relationship between entities may have a significant impact on user choices. For example, in restaurant recommendations, users are more concerned about the restaurant's geographic location information or the restaurant's star rating or price at the current time. In e-commerce, click behavior in recent sessions can also provide a strong signal for the user's next purchase. Therefore, context information and knowledge graphs can be effectively integrated into the recommendation system to improve the recommendation performance of the recommendation system.

[0103] Different from the related technical solutions of the first and second aspects, Figure 2 As shown, the embodiment of the present application provides another technical solution. Figure 2As shown, the embodiment of the present application provides a display method, which mainly includes the following steps:

[0104] Step S110 , the device displays the user interface 10 , and the user interface 10 displays: display content 101 .

[0105] Optionally, the display content 101 displayed by the user interface 10 may be a recommendation result obtained based on a recommendation algorithm, or may be a randomly recommended item. For example, during a cold start, the user interface 10 may randomly recommend the display content 101 .

[0106] In step S120, after the device detects the operation 102 acting on the displayed content 101, the user interface 20 is displayed. The user interface 20 displays: the displayed content 201. The displayed content 201 is related to the relational intent and the contextual intent. Among them, the displayed content 201 is the result of the recommendation model making a recommendation based on at least the embedded representation information. The embedded representation information is the representation information output by the graph neural network based on the relational intent and the contextual intent. The relational intent represents the user's attention to the relationship between the displayed content 101 and the associated entities, and the associated entities include entities with relationships in the knowledge graph of the displayed content 101. The contextual intent represents the user's attention to the context of the operation 102.

[0107] Optionally, the operation 102 acting on the display content 101 may be as follows: Figure 3 The clicks shown, such as favorite, add to cart, delete, can also be like, praise, comment and other operations. Display content 102 can be any content, such as text, numbers, symbols, icons, etc., or any combination thereof. For example, images, videos, item names, etc. Exemplarily, display content 101 can be a link or mirror corresponding to the display content exposed to the user. When the user clicks on the link or mirror, the device displays to the user the detailed content related to the display content 101 corresponding to the link or mirror.

[0108] Optionally, the associated entity of the display content 101 may be Figure 3 The provision of outdoor seats 105 shown in the figure may also include celebrity endorsements, hotel front desk services, music, etc. For example, when the user performs operation 102 on hotel 1, since hotel 1 provides outdoor seats 105, it can be considered that the user's relationship intention is: preferring hotels that provide outdoor seats 105. Therefore, if Figure 4 As shown, when making recommendations to users, hotels that provide outdoor seating 202 are given priority.

[0109] Optionally, the context of operation 102 may be time information 103, location information 104, etc. For example, when the user collects Hotel 1, the location is Beijing, and the time is Sunday. The user's contextual intention is: on Sunday, the user tends to choose a hotel in Beijing. Therefore, if Figure 4 As shown, when the time information 203 is Sunday and the user's location information 204 is Beijing, recommendations can be made to the user based on the information of Hotel 1.

[0110] It can be further understood that the user's contextual intention can be determined based on the user's operation 102 on multiple hotels, and obtained in combination with the corresponding context. The user's relational intention can be determined based on the user's operation 102 on multiple hotels, and obtained in combination with the associated entities of the multiple hotels. In other words, when the user has more interaction data, the tendency of the user's contextual intention and relational intention is more obvious and more accurate.

[0111] In a possible implementation, after the device displays the user interface 20, the display method further includes: after the device detects the second operation, displaying a user interface 30. The user interface 30 displays: at least one associated entity 303, the user's attention level 304 to the associated entity; and / or at least one context 301 of the operation 102, the user's attention level 302 to the context of the operation 102.

[0112] Optionally, the second operation may be sliding the user interface 20 to the left or clicking an icon “Interest Graph” or the like.

[0113] For example, Figure 5a As shown, the context 301 of operation 102 can be Beijing and Sunday, and the user's attention level 302 to the context of operation 102 is 0.425 and 0.215 respectively. The associated entity 303 can be providing outdoor seats, celebrity endorsement, providing music, and providing hotel front desk service; the user's attention level 304 to the associated entity is 0.312, 0.201, 0.102, and 0.051 respectively.

[0114] Optionally, the user can click or long press a related entity, context or corresponding attention level to modify or delete it to affect the recommendation results.

[0115] In one example, if Figure 5b As shown in the figure, users can also click the search control on the side of the "Interest Graph" to query a more detailed interest graph. Figure 5b The right side view is shown.

[0116] like Figure 5bAs shown, optionally, when the system recommends items to users, it can clearly reflect which information of the items the users pay more attention to in certain contexts. The degree of attention will be shown in the interface in the form of weights. In other words, the user interface (UI) is highly interpretable, which improves the user experience.

[0117] Optionally, the solution of the embodiment of the present application can be implemented on the client side, so that user data does not enter the cloud system to protect user privacy as much as possible. Therefore, the recommendation solution can be personalized, and real-time recommendation on the client side can be achieved based on the user's personalized recommendation model.

[0118] Optionally, since the graph structure of the embodiment of the present application has good generalization, different user behaviors, or cross-domain behaviors, or cross-end behaviors can be combined to achieve the same experience for users across domains or across ends.

[0119] In one example, if Figure 5b As shown, an exemplary description of the interest graph is given. For example, in a hotel recommendation scenario, in a certain interaction event of the user, the interaction context information includes location information, time information, and historical information. Among them, the location information is, for example, the city c1. The time information is, for example, the year c2, month c3, and week c4 when the interaction event occurred. Historical information is, for example, some information of the user's previous interaction event, which is used as reference information for the analysis of this interaction event. Therefore, the contextual intent characterizes the user's attention to different contextual information. In other words, the contextual intent characterizes the user's attention to the city where the hotel is located, or the user's attention to the day of the week, month, or year. For example, as Figure 5b As shown, the user's contextual intent shows that the user is most interested in city c1, with a score of 0.425.

[0120] like Figure 5bAs shown, for example, the shared intention characterizes the degree of association of the relationship with the entity. For example, the shared intention p1 includes the relationship r27 with the entity (music) with a degree of association of 0.515, and the relationship r22 with the entity (hotel front desk service) with a degree of association of 0.163. The shared intention p2 includes the relationship r35 with the entity (the cost of bringing your own wine) with a degree of association of 0.303, and the relationship r5 with the entity (outdoor seating) with a degree of association of 0.129. The shared intention p3 includes the relationship r22 with the entity (hotel front desk service) with a degree of association of 0.316, and the relationship r20 with the entity (star) with a degree of association of 0.272. The shared intention p4 includes the relationship r5 with the entity (outdoor seating) with a degree of association of 0.254, and the relationship r28 with the entity (smoking) with a degree of association of 0.166. The shared intention is shared by multiple users. For example, if a user is more concerned about outdoor seating, the weights of the shared intention p4 and the shared intention p2 in the user's relationship intention will be larger. For example, Fig.18 As shown, for example, user u114 is more concerned about the hotel front desk service. Therefore, among the relationship intentions of this user, the weight of the sharing intention p3 is the largest, which is 0.863, followed by the sharing intention p1, which has a weight of 0.102.

[0121] In another possible implementation, after the device displays the user interface 20, the display method further includes: after the device detects the third operation, displaying the user interface 40. The user interface 40 displays: at least one display content 101, and / or at least one associated entity, and / or at least one context of the operation 102.

[0122] Optionally, the third operation may be sliding the user interface 20 to the right or clicking an icon “interaction history” or the like.

[0123] For example, Figure 6 As shown, the displayed content 101 is hotel 1, the associated entity is outdoor chair 105, and the context of operation 102 is time, the location of the user during interaction, etc.

[0124] Optionally, the user can browse different interaction histories by clicking the "Previous" or "Next" buttons. The user can also delete an interaction event, and the deleted user event will not be used as the basis for user recommendation.

[0125] In combination with the display method of the embodiment of the present application, the embodiment of the present application also provides a recommendation method. For the sake of brevity of description, the implementation method of the display method of the embodiment of the present application is the same as the implementation method of the recommendation method, and will not be repeated.

[0126] like Figure 7 As shown, the embodiment of the present application provides a recommendation method, which mainly includes the following steps:

[0127] Step S210, based on the user's historical interaction data, determine the user's relational intent and contextual intent. The historical interaction data includes the user's first operation on the first displayed content and the context of the first operation. The relational intent represents the user's attention to the relationship between the first displayed content and the associated entity, and the associated entity includes the entity with a relationship in the knowledge graph of the first displayed content. The contextual intent represents the user's attention to the context of the first operation.

[0128] Step S220, based on the user's relational intention and contextual intention, the embedded representation information is obtained through the graph neural network.

[0129] Step S230: obtaining second display content through a recommendation model at least according to the embedded representation information.

[0130] In the embodiment of the present application, the first display content may be display content 101 , the first operation may be operation 102 , and the second display content may be display content 201 .

[0131] In step S210, the process of determining the relational intent and the contextual intent is described. The following is an exemplary description of the process of determining the relational intent and the contextual intent.

[0132] First, as Figure 8 As shown, an embodiment of the present application provides a combination of a personal graph and a knowledge graph to simultaneously consider the user's relational intent and contextual intent.

[0133] A personal graph refers to a graph that represents the context of a user's interaction. Exemplarily, a personal graph includes user nodes, item nodes, and contextual edges. User nodes are used to represent user information. Item nodes are used to represent items that need to be inferred user preferences. Contextual edges are used to represent information about the context of the user's interaction with an item.

[0134] like Figure 8 As shown, the embodiment of the present application embeds the knowledge graph into the personal graph, and obtains a combination graph of the personal graph and the knowledge graph. Figure 8 As shown in the figure, in one interaction event, user u1 interacted with item i1, item i2, and item i3 respectively. The context information of the user's interaction with item i1, item i2, and item i3 are C1, C2, and C3 respectively. In the knowledge graph, item i1, item i2 and entity v1 have a relationship r1. Item i1, item i3 and entity v2 have a relationship r2. Item i1, item i3 and entity v2 have a relationship r2, and entity v2 and entity v3 have a relationship r3.

[0135] In the knowledge graph on the item side, there may be relationships between items, and the related items can be treated as related entities. For example, Figure 8 As shown, there is a relationship between item i1 and item i2, and there is a relationship between item i2 and item i3. When processing the knowledge graph information of item i1, the associated item i2 can be processed as a general associated entity, and the relationship between item i1 and item i2 can be processed as a general relationship with an associated entity.

[0136] like Figure 8 As shown in the combined graph of user u1, the personal graph part is T-pg, including the nodes within the dotted line range as shown in Figure 5. The knowledge graph part is T-kg, including Figure 8 Nodes within the dashed line are shown.

[0137] It can be further understood that if Figure 8 As shown, in the graph neural network layer h1, there may also be other user's combined graphs, which will not be described in detail in the present embodiment. The graph neural network can be multi-layered, such as Figure 8 The network layer h1, network layer h2, network layer h3, etc. are shown.

[0138] Direct application in CARS Figure 8 The combination of the personal graph and the knowledge graph shown is very difficult for the following reasons:

[0139] First, the existence of a large number of heterogeneous nodes means that unreasonable convolution methods will bring too much noise in information propagation. A large number of heterogeneous nodes, including the connection edges in the graph structure, such as context connection edges, relationship connection edges, etc.

[0140] Second, the connection between entities and user information and context information is weak, which may also weaken the extent to which user information and context information benefit from the knowledge graph. Figure 8 As shown in the figure, the information of the entity node is transmitted to the item node, and the information of the item node is transmitted to the context information node and the user node through the reverse process of the attention mechanism. Therefore, there are many transmission layers between the entity and the user information and the context information, and the connection is weak. In other words, in the detailed representation results of user information and context information, the influence of item attributes is relatively low, which weakens the degree to which user information and context information benefit from the knowledge graph.

[0141] In the embodiment of the present application, a graph structure of a pairwise intent graph (PING) is proposed for context-aware recommendation. The context information and item attributes are integrated into the fusion intent node, and when the knowledge graph on the item side is introduced, a large number of heterogeneous nodes are avoided, the computational complexity is reduced, and excessive noise is avoided.

[0142] In an embodiment of the present application, by fusing the relational intent representing the attributes of an item and the contextual intent representing the contextual information, the association between the semantic information of the item and the user information and the contextual information is enhanced, and the transmission steps between the attributes of the item and the user information and the contextual information are reduced, so that the user information and contextual information after refined representation can represent the relevant information in the knowledge graph of the item.

[0143] In the embodiment of the present application, a concept of a pairwise intent node is introduced to model the user's preference for each interaction. A pairwise intent graph can be constructed

[0144]

[0145] In formula (1), It is used to represent the set of users, items, and objects, and is represented as a node in the PING graph. Used to represent the relationship between users and items The relationship between items Relationships between items and entities Relationships between entities The collection of nodes is represented as connecting edges in the PING graph.

[0146] Furthermore, all historical interactions of a user in CARS can usually be organized into a personal graph. The personal graph is a related representation centered on the user node representing the user. As shown in Figure 5, the nodes include users and interacted items, the edges include the interactions between users and items, and the temporal relationships between items, i.e. and The list of context features is used as After mapping items and entities, the personal graph can be further integrated with the knowledge graph KG. In other words, the mapping relationship between users and interacted items can be obtained from the personal graph; the corresponding mapping relationship between items and entities can be obtained from the knowledge graph. Therefore, the paired intent graph is the result of the combination of the personal graph and the knowledge graph.

[0147] Further, similarly, the mapping relationship r between items and entities can be viewed as and The edge attributes on .

[0148] For example, Figure 8 As shown, the user u1 in the user set interacts with the items i1, i2, and i3 in the item set. From the relationship intention attention, it can be seen that, for example, the item i1 has relationships r1 and r2 with the physical objects v1 and v2 in the physical object set, respectively.

[0149] In the embodiment of the present application, In the PING graph, it is represented as a node. In the PING graph, it is represented as a connection edge. It is represented as a fusion intention node and embedded into the PING graph, so that the PING graph is represented as the connection between nodes.

[0150] Furthermore, the fusion intent node represents the intent fused by the relational intent and the contextual intent. That is to say, in the process of updating the features of the graph neural network, the fusion intent of the relational intent and the contextual intent can be propagated to other nodes through the fusion intent node to update the features of other nodes.

[0151] Optionally, relationship intent nodes and context intent nodes may also be set in the PING graph, and relationship intent and context intent may be propagated to other nodes through the relationship intent nodes and context intent nodes, respectively, to perform feature updates on other nodes.

[0152] Secondly, an exemplary explanation is given of the relational intent driven by the semantic information of the knowledge graph.

[0153] Alternatively, as Fig. 9 As shown, a set of shared intents P = {p1, p2, ...} is defined. Shared intent P, also known as relational sub-intent, can be derived from the relation set in the knowledge graph. Exemplarily, an algorithm for shared intent P is shown in formula (2).

[0154]

[0155] In formula (2), e p An expression representing the sharing intention P, e r An expression representing a relationship in a knowledge graph, e r is the embedding of relation r. rp It represents the degree of attention of intention p to relation r. R represents the set of relations in the knowledge graph. P represents a relational intention in the shared intention P, that is, p∈P.

[0156] For example, α rp One way to calculate is shown in formula (3).

[0157]

[0158] In formula (3), w rp The weight corresponding to the relationship intention p and the relationship r is a trainable weight of the degree of association between a relationship r and a shared intention P. r′p Represents the weight corresponding to the relationship intention p and the relationship r′.

[0159] That is to say, the shared intent P and all the relations in the knowledge graph can be processed by the attention mechanism and then summed. Therefore, the shared intent P represents the degree of attention of each intent p to different relations r.

[0160] Furthermore, since the sharing intention P is global, the embodiment of the present application further introduces user information to obtain personalized relationship intention t′ u That is, the relational intention t′ u User information is also characterized.

[0161]

[0162] In formula (4), Indicates relational intention t′ u The expression of β up is to obtain the relationship intention t′ u The parameter is used to measure the weights of different users for different sharing intentions P.

[0163] For example, β up One way to calculate is shown in formula (5).

[0164]

[0165] In formula (5), e u Used to represent user information. up Characterizes the influence of user information on sharing intention P.

[0166] like Fig. 9 As shown in the figure, it shows the connection relationship between user nodes, relationship intent nodes, virtual nodes of shared intent, and relationship connection edges. Shared intent is also called relationship sub-intent.

[0167] like Fig. 9As shown in Figure 1, the shared intent P includes shared sub-intentions p1 and p2. The shared intent P is an attention mechanism for the degree of association between items and item attributes. Item attributes can be understood as the relationship between items and associated entities, such as Fig. 9 The relationships r1, r2, and r3 are shown.

[0168] For example, in an application scenario where the item is a restaurant, the shared sub-intent p1 can represent that there is a strong correlation between the geographical locations of restaurants, and the shared sub-intent p2 can represent that there is a strong correlation between the types of dishes in restaurants.

[0169] like Fig. 9 As shown, further, the relationship intention t′ u The calculation process of can also be understood as the result of an attention mechanism processing the shared intention P. For example, Fig. 9 As shown, the relationship intention t′ u It may include relational sub-intention tr1, relational sub-intention tr2, and relational sub-intention tr3. Relational sub-intention tr1 represents the preference of user u1 for item i1, relational sub-intention tr2 represents the preference of user u1 for item i2, and relational sub-intention tr3 represents the preference of user u1 for item i3. Optionally, in the calculation process, relational intention t′ u It can be a set or vector including multiple relationship sub-intents.

[0170] For example, in an application scenario where the item is a restaurant, the user is more concerned about the restaurant's geographical location. In the relationship sub-intention tr1, the attention weight of the shared sub-intention p1 is larger, and the attention weight of the shared sub-intention p2 is smaller.

[0171] Furthermore, in order to characterize that the sharing intention P is shared among multiple users. Fig. 9 As shown in the figure, when user u2 interacts with an item, there is a relational intent, which can be understood as an attention mechanism for the shared intent P. For example, when user u2 interacts with items i1, i2, and i3, there are three relational sub-intents. The relational sub-intent is the result of the attention mechanism processing of the shared sub-intent p1 and the shared sub-intent p2.

[0172] Thirdly, the contextual intent driven by contextual information is exemplified.

[0173] For example, assuming that the number of context features associated with each object is defined as Z, the vector of context features is Since the user's behavior may also be affected by previous behavior, not just the context, the present embodiment also introduces a context-specific feature To represent the last interactive item before the current instance, that is, the features associated with the item of the last interaction. Then, based on formula (6) and formula (7), the user's contextual intention for the object can be obtained.

[0174]

[0175] In formula (6), W1,W2∈R d×d , is a trainable parameter and d is the size of the contextual feature embedding convolutional model. Indicates the characteristics associated with the user. Represents contextual features. Represents the attention weight of the z-th context feature for each context intent.

[0176]

[0177] In formula (7), is the embedded representation of the intention node corresponding to user u and the kth physical object. In a context scenario, when a user interacts with the kth interaction event, if the user pays high attention to a feature in the context feature, the weight value of the feature in the context feature is It will be larger and will be more emphasized when the information is disseminated.

[0178] Optionally, based on each user interaction record, combined with the characteristics associated with the user Contextual Features Features associated with the last item interacted with Calculate the user's attention to different context features Based on context features, Aggregate into the user’s contextual intention for the kth object.

[0179] like Fig.10 As shown, the embodiment of the present application constructs a contextual intention t representing the user to the item by using an attention mechanism for contextual features. k . Exemplary, contextual intent kIt may include context sub-intent tc1, context sub-intent tc2, and context sub-intent tc3. Context sub-intent tc1 represents the degree of attention of user u1 to context feature c(1,1), context feature c(1,2), and context feature c(1,2) of item i1 interaction; context sub-intent tc2 represents the degree of attention of user u1 to context feature c(2,1), context feature c(2,2), and context feature c(2,2) of item i2 interaction; context sub-intent tc3 represents the degree of attention of user u1 to context feature c(3,1), context feature c(3,2), and context feature c(3,2) of item i3 interaction. Optionally, in the calculation process, context intent t k It can be a set or vector including multiple relationship sub-intents.

[0180] Fourthly, optionally, in order to construct a paired intent graph, the present application embodiment proposes a user intent fusion (UIF) solution. Further, the fused user intent can capture the fine-grained user intent in each item to embed the corresponding fused intent node. In other words, user intent fusion UIF is Fig. 9 The relationship shown is intended and Fig.10 The fusion of contextual intent shown in the figure is used to mine the fine-grained user intent in the process of user interaction with each item. Fine-grained user intent can be understood as the fusion of relational intent and contextual intent. The fused user intent can simultaneously represent the information of relational intent and contextual intent. It is a more fine-grained user intent and can more accurately represent and distinguish user intent.

[0181] Optionally, the relationship intent t′ u Contextual Intent k The purpose of fusing user intent is to obtain the relational intent t′ through independent modeling. u Contextual Intent k , to avoid introducing noise. The context information is the connection edge between the user node and the item node. k Represents the user's attention mechanism to contextual information. Relationships are the edges connecting item nodes and entity nodes. Relationship intent t′ u It represents the user’s attention mechanism on the relationship between items and entities. Therefore, the relationship intention t′ is obtained by independent modeling u Contextual Intent k , reducing the number of connecting edges and avoiding the introduction of noise in the process of adding knowledge graph information.

[0182] Optionally, the relationship intent t′ uContextual Intent k The fusion of t′ enhances the association between semantic information and user information and context information. In other words, by fusing intent nodes and propagating information, user information and context information can represent semantic information more accurately. u Contextual Intent k The fusion of entities always keeps a strong connection with user information and context information. In other words, the fused fusion intent is embedded in the fusion intent node, and the fusion intent node is the central hub of the graph representation corresponding to the user of the graph neural network. Therefore, as the number of times the recommendation model is learned increases, the connection between the entity and the user information or context information will not be weakened.

[0183] For example, the process of user intent fusion can use the linear fusion method to perform weighted summation on two sub-intents. For example, mean fusion can be expressed as

[0184] In other examples, the process of user intent fusion can use variable weights to t′ u Contextual Intent k For weighting, you can also use nonlinear fusion methods, such as ReLU function, sigmoid function, etc.

[0185] like Fig.11 As shown, the user's relationship intention to the item t′ u Contextual Intent k By fusion, we can get the user's fusion intention for the item Exemplarily, the relationship intention tr1 and context intention tc1 of user u1 for item i1 are fused to obtain the fused intention tf1 of user u1 for item i1. The relationship intention tr2 and context intention tc2 of user u1 for item i2 are fused to obtain the fused intention tf2 of user u1 for item i2. The relationship intention tr3 and context intention tc3 of user u1 for item i3 are fused to obtain the fused intention tf3 of user u1 for item i3.

[0186] In step S220, it is defined that the user's relational intent and contextual intent are taken into account during the feature update process of the graph neural network. In other words, the graph structure of the input graph neural network and the information propagation process between nodes are designed so that the features of the nodes in the graph structure output by the graph neural network represent the user's relational intent and contextual intent. The following is an exemplary description of the graph structure of the input graph neural network and the information propagation process between nodes.

[0187] On the one hand, the embodiment of the present application proposes a PING graph structure to embed relational intent and contextual intent. The PING graph structure is a graph representation of a graph neural network corresponding to a user. The graph representation is input into the graph neural network, and the representation information of the nodes of the graph representation corresponding to the user is obtained based on the information propagation of the nodes between the network layers of the graph neural network.

[0188] Alternatively, if Fig.12 As shown, contextual intent nodes and relationship intent nodes can be set between user nodes and item nodes in the PING graph structure to update the features of user nodes and item nodes through the graph neural network. Fusion intent nodes can also be set between user nodes and item nodes in the PING graph structure to update the features of user nodes and item nodes through the graph neural network. Among them, the PING graph structure with fusion intent nodes is more concise, and the graph neural network is more efficient in updating the features of the PING graph.

[0189] For example, Fig.12 As shown in FIG. 1 , the PING graph structure is a graph representation in which a fusion intention node is embedded between the user node and the item node to form a user-fusion intention-item corresponding graph representation.

[0190] In one example, if Fig.12 As shown in Figure 1, the PING graph structure includes user nodes, item nodes, intent nodes, and connection edges. The connection edges are used to connect user nodes and intent nodes, and intent nodes and item nodes. Fig.12 As shown in the figure, the relational intent and contextual intent are processed and the processed results are embedded in the fusion intent node. With the fusion intent node as the hub, the message is propagated to the user node and the item node to obtain accurate user representation and item representation. The relational intent and contextual intent can also be propagated independently to other nodes through the intent node.

[0191] Alternatively, if Fig.13As shown, the PING graph structure may also include user nodes, item nodes, fusion intent nodes, context intent nodes, context nodes, and connection edges. The connection edges are used to connect user nodes and fusion intent nodes, user nodes and context intent nodes, context intent nodes and context nodes, fusion intent nodes and item nodes, and item nodes and context intent nodes. User nodes represent users, item nodes represent items, fusion intent nodes represent user fusion intents, context intent nodes represent user context intents, and context nodes represent context information of user interactions. Fusion intent includes the fusion result of context intent and relationship intent. Context intent represents the user's attention to the context information of the interaction. Relationship intent represents the user's attention to the relationship between items and associated entities. Associated entities include entities in the knowledge graph that have a relationship with items.

[0192] like Fig.14 As shown, optionally, the PING graph structure also includes entity nodes and relationship connection edges. The relationship connection edges are used to connect item nodes and entity nodes. The entity nodes represent the information of the associated entities, and the relationship connection edges represent the relationship between the items and the associated entities.

[0193] like Fig.15 As shown, it can be further understood that the number of users in the graph representation corresponding to the user is, for example, N, where N is an integer greater than zero. The number of users N can be the number of all users or the number of active users. Each user node can be connected to one or more fusion intent nodes. Each fusion intent node can be connected to one or more item nodes. A fusion intent node can be connected to multiple item nodes, indicating that the user has the same intention for the items of the multiple item nodes. Fig.15 In order to simplify the representation, some connection edges are not shown, including the connection edges between the context intent node and the user node and the item node. In other words, the PING graph structure is a graph structure containing N users and the interaction events of these N users. Therefore, when a new user or a new interaction event appears, it is necessary to reconstruct the graph and learn and train the graph neural network.

[0194] like Fig.16 As shown, the graph representation corresponding to the user is input into the graph neural network, and the graph neural network updates the features of the nodes or connecting edges of the graph representation corresponding to the user to obtain the node representation information of the graph representation corresponding to the user. The node representation information of a node in the graph representation corresponding to the user represents the attribute information of the node and the information of the adjacent nodes. The adjacent nodes can be first-order neighbor nodes or high-order neighbor nodes. Exemplarily, the node representation information can represent the node information of the graph representation corresponding to the user in the form of a vector.

[0195] The workflow of the embodiment of the present application is exemplarily described. From the user's behavior log, etc., historical interaction data of multiple users are obtained, and the historical interaction data includes user information, item information, and interactive context information. Based on the item information, the knowledge graph of the interactive item is obtained. Based on the relationship between the interactive item and the associated entity, the relationship intention is calculated. The associated entity is an entity related to the interactive item in the knowledge graph. Based on the interactive context information (interactive context information) in the interactive data, the context intention is calculated. The relationship intention and the context intention are fused to obtain the fusion intention. The graph structure of PING is input into the graph neural network in the form of a graph representation. The initial features of the user node, item node, fusion intention node, context intention node, and context node of the graph structure of PING are user information, item information, fusion intention, context intention, and context information, respectively. The nodes of the graph structure of PING are updated with features through the graph neural network to obtain a representation data set. The representation data set includes embedded representations of user nodes, item nodes, fusion intention nodes, context intention nodes, and context nodes.

[0196] In the second aspect, an exemplary description is given of information propagation in a graph neural network. In the embodiment of the present application, with the help of a knowledge graph, users can be effectively helped to capture item attributes and context information.

[0197] like Fig.17 As shown, an embodiment of the present application provides a schematic diagram of user information dissemination, using a fusion intent node as a hub to complete the required information dissemination.

[0198]

[0199] In formula (8), Represents the user information in the kth interaction event passed to the hth layer of the graph convolution, which is embedded in the user node. Represents the item information embedded in the item node at the h-1 layer, where represents the fusion intent embedded in the fusion intent node at the h-1 layer, where

[0200] Optionally, in formula (8), the information of the item nodes at the h-1 layer and the information of the fusion intention nodes are transmitted to the user nodes at the h-th layer through the information propagation of the second-order neighbors. Alternatively, the information of the item nodes at the h-2 layer is transmitted to the fusion intention nodes at the h-1 layer through the information propagation of the first-order neighbors, and then the information of the fusion intention nodes at the h-1 layer is transmitted to the user nodes at the h-th layer.

[0201] Optionally, the fusion intent includes the fusion result of the relational intent and the contextual intent. The relational intent represents the user's attention to the relationship between the item and the associated entity, and the contextual intent represents the user's attention to the contextual information. Therefore, the relational information embedded in the relational connection edge and the contextual information embedded in the context node are propagated to the fusion intent node in the h-1 layer through the first-order neighbor or the multi-order neighbor to update the fusion intent embedded in the fusion intent node.

[0202] Optionally, as shown in formula (9), based on different interaction events, an aggregation function is used to obtain the embedding of each user at layer h.

[0203]

[0204] Alternatively, as shown in formula (10), refined user embeddings can be obtained by averaging the embeddings of each layer.

[0205]

[0206] like Fig.18 As shown, an embodiment of the present application provides a schematic diagram of item information propagation. Using paired intent nodes as the center, the items can effectively capture user and context information with the help of knowledge graphs.

[0207]

[0208] In formula (11), Represents the information passed to item i associated with the kth interaction event in the hth layer of the graph convolution. Represents the embedded information of the user node at the h-1th layer.

[0209] Optionally, in formula (11), the information of the user nodes and the fusion intention nodes of the h-1 layer are transmitted to the item nodes of the h-th layer through the information propagation of the second-order neighbors. Alternatively, the information of the user nodes of the h-2 layer can be transmitted to the fusion intention nodes of the h-1 layer through the information propagation of the first-order neighbors, and then the information of the fusion intention nodes of the h-1 layer can be transmitted to the item nodes of the h-th layer.

[0210] Optionally, the fusion intent includes the fusion result of the relational intent and the contextual intent. The relational intent represents the user's attention to the relationship between the item and the associated entity, and the contextual intent represents the user's attention to the contextual information. Therefore, the relational information embedded in the relational connection edge and the contextual information embedded in the context node are propagated to the fusion intent node in the h-1 layer through the first-order neighbor or the multi-order neighbor to update the fusion intent embedded in the fusion intent node.

[0211] Alternatively, as shown in formula (12), the embedding of item i at layer h can be obtained by using an aggregation function and receiving additional information. There are k interaction events for the user's interaction with item i, so it is necessary to use an aggregation function to aggregate the k events to obtain the embedding representation of item i.

[0212] That is to say, when Fig. 9 When the PING graph structure shown in is input into the graph neural network, as shown in formula (12), the embedding of item i in layer h can be obtained by using the aggregation function. Fig.10 When the PING graph structure shown is input into the graph neural network, as shown in formula (12), the embedding of item i in layer h can be obtained by using the aggregation function and receiving additional information. The additional information represents the information of the entity node and the information of the relationship connection edge transmitted to the item node.

[0213] Exemplarily, as shown in formula (12), the entity node information and the relationship edge information are subjected to XOR operations and then averaged to obtain the attachment information.

[0214]

[0215] In formula (12), represents the set of relations and entities associated with item i, represents the number of sets, represents the entity embedding at the h-1th layer of the graph convolution, where

[0216] As shown in formula (13), a refined embedding representation of the item node can be obtained by averaging the embeddings of each layer.

[0217]

[0218] like Fig.19 As shown, an embodiment of the present application provides a schematic diagram of context information propagation. For context features, as shown in formula (14), user and item information will first be propagated to the context intent node. As shown in formula (15), the context intent node then propagates information to different context nodes, and different context features will distribute information according to the context attention distribution.

[0219]

[0220] In formula (14), Represents the information passed to the contextual intent node associated with the kth interaction event in the hth layer of the graph convolution.

[0221]

[0222] In formula (15), represents the information representation of the context information associated with the kth interaction event in the hth layer of the graph convolution. Since the information propagation process of formula (15) is equivalent to the inverse process shown in formula (7), it is necessary to embed the context intent node conduct The weight of .

[0223] Optionally, as shown in formula (16), by aggregating different interaction events, the embedding of the context information of the hth layer can be obtained.

[0224]

[0225] Optionally, as shown in formula (17), the final refined embedding can be obtained.

[0226]

[0227] In the embodiment of the present application, since the context feature is only related to the context intent, the information propagation of the context feature only uses the sub-intent node t obtained by the context intent attention. k , rather than pairs of intent nodes.

[0228] In addition, since the embedded representation of the user node and the embedded representation of the item node in the next layer require the contextual intention node of the h layer to transmit information, it is necessary to feedback information to the contextual intention node of the h layer. Exemplarily, the calculation formula can be shown as formula (18).

[0229]

[0230] like Fig. 20 As shown in the figure, when the graph representation of the input graph neural network contains entity nodes, the information of the entity nodes can also be transmitted and updated. For example, the graph representation is as follows Fig.10 The PING graph structure shown. The embodiment of the present application provides a schematic diagram of information propagation to an entity node. As shown in formula (19), exemplary, the information propagation of an entity node involves information transmission between an item node and a relationship connection edge. In other examples, the information of a high-order neighbor of an entity node can also be propagated to the entity node, such as the information of a second-order neighbor.

[0231] For example, in formula (19), the information of the item nodes at the h-1 layer and the information of the relationship connection edges are subjected to the same or same operation and then averaged to obtain the embedded information of the entity nodes at the h layer.

[0232]

[0233] In formula (19), Represents a collection of relationships and items associated with entity v. Indicates the number of sets.

[0234] Therefore, as shown in formula (20), after performing the pairwise intent joint graph convolution, a set of corresponding embedding representations can be obtained, including the embedding representation of the user node, the embedding representation of the item node, and the embedding representation of the interactive context node. The embedding representation is a refined embedding vector. By aggregating multiple interaction events of multiple users, a representation dataset can be obtained.

[0235]

[0236] Optionally, the characterization data set may be updated according to certain requirements, such as updating the characterization data set in a fixed event segment or by conditional triggering.

[0237] Furthermore, when updating the representation dataset, the PING graph structure needs to be rebuilt to add or remove new users or new interaction events. The rebuilt PING graph structure is trained through the graph neural network to obtain an updated representation dataset. The updated representation dataset includes the updated embedded representation of user nodes, the embedded representation of item nodes, and the embedded representation of interactive context nodes, which are input into the downstream recommendation subnetwork, and the recommendation subnetwork outputs the recommendation results.

[0238] In the embodiment of the present application, relational intent and contextual intent are fused by introducing relational intent. That is, the user, item, all contexts of interaction with the item, and the relationship between the item and the entity are explicitly modeled, so as to deepen the understanding and learning of the user's interaction intent, and more refined user intent is represented in the fused intent.

[0239] In other words, the corresponding relationship connection edges between item nodes and entity nodes, and the contextual information of user-item interactions are modeled to obtain the PING graph structure, and through information transmission, finely characterized users and interaction events are obtained. In this way, not only the semantic information and contextual features of the knowledge graph are integrated into the model, but also the user information, item information, and contextual information can be continuously refined in the process of information propagation of the graph neural network by integrating the intent nodes. The accuracy of fine-tuning can be improved by increasing the number of layers of the neural network or increasing the number of iterations of the graph neural network.

[0240] It can be further understood that during the training process of the graph neural network, the first user label is used as the theoretical value for training, and the loss function is calculated based on the output value of the graph neural network and the theoretical value. Based on the calculation result of the loss function, adjust w in formula (3)rp Trainable weights or trainable parameters such as W0, W1, W2, b1 in formula (6) can be iteratively updated. Optionally, based on the calculation result of the loss function, other parameters in the graph neural network can also be iteratively updated, such as the bias value of the network layer, the weight value of the graph neuron, etc.

[0241] In an embodiment of the present application, a graph convolution representation method based on PING uses a fusion intent node as an intermediate node (hub) to transmit the knowledge graph information and item information on the item side to the user node respectively, and learn a more refined embedding representation of the user node.

[0242] In the embodiment of the present application, the contextual intent node is used as the hub, and the item information and user information are passed to the context node, so that more detailed context information can be obtained. By fully learning the context information, user events can be more accurately segmented. By integrating the relationship intent and the contextual intent, the context information and semantic information can be fully learned, and the user events can be more accurately segmented to achieve the purpose of improving the recommendation effect. User events refer to user interaction events, and interaction events further include user information, item information, and context information of the user's interaction with the item.

[0243] In the embodiment of the present application, the fusion intention node is used as the hub, and the item information, context information and relationship information are passed to the fusion intention node to obtain fine context information. In other words, the embedded representation of the user node, the embedded representation of the item node and the embedded representation of the context node are fully refined through the context information and the knowledge graph information, so that the user event can be segmented more accurately.

[0244] Furthermore, in the convolutional graph, by fusing the intent nodes and refining the information of the knowledge graph and the context information, the representation of users and items is further refined, making the representation of users and items more accurate. In other words, by propagating information on the entity nodes and context nodes, the information of the entity nodes and context nodes is refined. The refined information of the entity nodes and context nodes is then transmitted to the user nodes and item nodes through information propagation, thereby further refining the embedded representation of the user nodes and item nodes.

[0245] In step S230, a process of generating a recommendation result is described. At least according to the embedded representation information, the second display content is obtained through the recommendation model.

[0246] like Fig.21As shown, the embodiment of the present application provides a downstream recommendation model. That is to say, the embodiment of the present application can be understood as a recommendation system that completes the recommendation task, and the downstream recommendation model is a part of the recommendation system, which realizes the recommendation calculation function of the recommendation task. Fig.13 As shown, the recommendation model includes an initial embedding layer, a fine embedding layer, a feature interaction layer, and an output layer. The initial embedding layer is used to embed the user's interaction data, including the user's information, the information of the interactive item, and the context information of the interaction. The fine embedding layer is used to organize the initial embedding layer. The feature interaction layer is used to perform feature interaction on the data of the fine embedding layer, and the output layer outputs the recommendation result. The recommendation result can be a list item or a single item.

[0247] In an embodiment of the present application, the data sorting process of the fine embedding layer is described. After the representation data set output by the graph neural network is embedded into the embedding layer of the downstream recommendation model, the information of the current or historical user, the information of the current or historical interacted items, and the context information of the current or historical interactions are obtained from the initial embedding layer. The corresponding embedded representation of the user, the embedded representation of the item, and the embedded representation of the context information are found from the representation data set. The embedded representation of the user, the embedded representation of the item, and the embedded representation of the context information are input into the feature interaction layer for feature interaction. In other words, the recommendation system can make real-time recommendations to the user based on the interaction events of the current user. Furthermore, the embedded representation of the user is also referred to as the user representation, the embedded representation of the item is also referred to as the user representation, and the embedded representation of the context information is also referred to as the context feature. Furthermore, the embedded representation information of the user includes the embedded representation of the user, the embedded representation of the item, and the embedded representation of the context information. The embedded representation information of the user refers to the information representation of the embedding layer or input layer corresponding to the user embedded in the downstream recommendation model or recommendation system.

[0248] In an embodiment of the present application, the downstream recommendation model may be matrix factorization (MF), factorization machines (FM), neural networks (NN), etc.

[0249] Furthermore, MF-based methods usually model the interactions between the embedded representations of users, items, and contextual information in the form of multiple matrices or tensors, and then solve them using coupled matrix decomposition and tensor decomposition, respectively.

[0250] Furthermore, FM-based methods replace matrix decomposition with the structure of factorization machines, which can capture the interactions between different features more effectively and accurately to improve high-order modeling and nonlinear relationships between features.

[0251] Furthermore, among the neural network-based methods, since most of the MF-based methods belong to shallow neural networks, the NN-based methods have the advantages of stronger model expression ability, more flexible model structure and better suitability to business scenarios by introducing machine learning or deep learning.

[0252] Furthermore, some complex network structures can be introduced, such as attention mechanism, convolutional network technology, etc.

[0253] In the embodiment of the present application, the graph neural network (GNN) may also be a graph convolution network (GCN). In other words, the graph convolution network may also implement the related functions of the graph neural network in the embodiment of the present application.

[0254] In the embodiment of the present application, by constructing pairs of intention nodes (relation intention t′ u Contextual Intent k ), the embodiment of the present application effectively utilizes the knowledge graph to enhance the high-order information collaboration between users, items and contexts, thereby further alleviating the data sparsity problem of node representation information and improving the recommendation performance.

[0255] Furthermore, due to the information propagation through multiple graph neural network layers, the node representation information output by the graph neural network can represent the information of high-order neighbors.

[0256] Furthermore, the representation information of the node output by the graph neural network is no longer a sparse matrix because it represents the information of the higher-order neighbors. That is to say, in the initial representation, the information of a node and its first-order neighbors is non-zero, the initial representation of the node and its higher-order neighbors is zero, and the initial representation matrix of the node is a sparse matrix. Therefore, when the node represents the information of the higher-order neighbors, the representation of the node and its higher-order neighbors is also non-zero, so the refined representation matrix of the node is a non-sparse matrix.

[0257] Furthermore, if Fig. 22As shown, an embodiment of the present application provides an architecture of a recommendation system. The recommendation system mainly includes an offline processing stage and an online processing stage. In an embodiment of the present application, a graph structure of a pairwise intent graph (PING) is constructed based on the knowledge graph and interaction context of the item. Based on the PING graph structure, the data of the offline feature library is processed to generate training samples. With the PING graph structure as the graph representation, the training samples are input into the nodes in the graph representation, and the nodes in the graph representation are updated with features through the graph neural network (GNN), and a representation data set of the nodes in the graph representation is output. The representation data set includes an embedded representation of the user, an embedded representation of the item, and an embedded representation of the context information. The representation data set is embedded into the embedding layer of the offline model. When the training sample is used to train the offline model, the embedded representation of the user corresponding to the user of the training sample, the embedded representation of the item, and the embedded representation of the context information are obtained from the representation data set, and the acquired embedded representation of the user, the embedded representation of the item, and the embedded representation of the context information are input into the offline model for learning and training. The PING graph structure includes the information of the knowledge graph of the item and the information of the user interaction context, so the embedding representation of the nodes in the output graph representation is more refined. In other words, the input data of the offline model more finely represents the contextual information of users, items, and interactions, so the training results of the offline model are more accurate. It can be further understood that more refined training samples can lead to more refined training results, and a more refined neural network model can be obtained after training.

[0258] Furthermore, a behavior wide table is extracted from the log generated by the user operation, and the log includes operations such as exposure and clicks. Based on the extracted behavior wide table, a first user tag is generated. The behavior wide table is a list of user behaviors, also known as a behavior log table. The first user tag is used to characterize the user's signal preference for an item and is a training tag for the training sample. For example, if a bit of the first user tag is 0 bit, it means that the user does not like the item corresponding to the bit; if a bit of the first user tag is 1 bit, it means that the user likes the item corresponding to the bit.

[0259] Further, a basic data table is generated from the source database. The source database is, for example, a user database, an item database, etc. Based on user behavior and the basic data table, feature mining is performed to obtain an offline feature library. The offline feature library refers to feature data related to the user, such as user features, item features, etc., which are training data of the training sample. In other words, the data of the offline feature library is processed into a form that conforms to the PING graph structure, and then input into the offline model to obtain detailed user information, item information, and context information, so as to obtain better recommendation results. The recommendation result is then compared with the first user tag, and based on the comparison result, the value of the trainable weight in the offline model is modified. After multiple iterations, a trained offline model is obtained. Among them, the data of the offline feature library is processed into a form that conforms to the PING graph structure, such as adding information about entities associated with items in the knowledge graph. When the PING graph structure of the data of the offline feature library needs to be displayed, the display method can be designed through the application program interface (API). The application program interface is a set of definitions, programs and protocols, and the mutual communication between computer software is realized through the API interface.

[0260] Optionally, in the online processing stage, the main work content includes: generating a second user tag based on the user request; generating a prediction sample based on the second user tag and the online feature library; the prediction sample includes user information, item information, and context information; inputting the prediction sample into the online model, obtaining the refined user information, item information, and context information corresponding to the prediction sample from the representation data set, and inputting them into the online model for prediction to obtain the recommendation result. Among them, the second user tag is a null value tag. That is to say, before the online model makes a prediction, the user's preference is unknown, so the second user tag is set to a null value.

[0261] Furthermore, the offline feature library in the offline processing stage is processed online. The offline model in the offline processing stage is processed online to obtain an online model. The online processing method is not described in detail.

[0262] Therefore, it can be seen from the offline processing stage and the online processing stage that the recommended method of the embodiment of the present application modifies the embedding layer of the offline model and the online model. In other words, the detailed user information, item information, and context information are embedded into the embedding layer of the offline model.

[0263] In the embodiment of the present application, a graph neural network is introduced in the process of refining user information, item information, and context information. The initial user information, item information, and context information are input into the graph neural network, and refined user information, item information, and context information are obtained through information propagation at the network layer.

[0264] In an embodiment of the present application, a fusion intent node is added to the node of the network layer. Furthermore, a relational intent is introduced through the knowledge graph. Contextual intent is introduced through the contextual information of user interaction. The relational intent and the contextual intent are fused to obtain a fusion intent. The fusion intent is embedded into the fusion intent node of the network layer. In other words, the information of the fusion intent node includes both the contextual information of the user's interaction with the item and the relationship between the item and the associated entity. The modeling process of the fusion intent node is an explicit modeling with strong interpretability and easy adjustment and modification of related parameters.

[0265] In the embodiment of the present application, the graph neural network integrates neighbor information through a message passing mechanism. The multi-layer stacked neighbor information allows nodes to access the information of high-order neighbors. For example, Fig.23 As shown in the figure, in the graph structure, node A is the root node, the first-order neighbors of node A are node B and node C, and the second-order neighbors of node A are node D and node E. During one information propagation, the information of the first-order neighbors is propagated. For example, when the information of the first layer is propagated to the second layer, node A receives the information of nodes B and C, node B receives the information of node D, and node C receives the information of node E. Therefore, when a three-layer graph neural network is used for information propagation, the information of nodes D and E can be propagated to node A. In other words, node A receives the information of the second-order neighbors, and the propagation process can be referred to as follows Figure 3 The information propagation process of the gray nodes is shown. Furthermore, when multiple network layers are stacked, nodes can access the information of high-order neighbors.

[0266] In the embodiment of the present application, since the context of the interaction is different when the user performs any interactive event, the user's intention is also different for each interactive action. The embodiment of the present application introduces fusion intent, adds fusion intent nodes, and models the knowledge graph and context information on the item side to infer the user's intention. The knowledge graph on the item side, also known as the item attributes of the item, includes the items interacted by the user, the entities associated with the item, the relationship between the item and the associated entity, etc.

[0267] Based on Figures 1 to 23 The same inventive concept, such as Fig.24 As shown, an embodiment of the present application provides a recommendation device 300, including: an intention determination module 310, which is used to determine the user's relationship intention and contextual intention based on the user's historical interaction data. The historical interaction data includes the user's first operation on the first displayed content and the context of the first operation. The relationship intention represents the user's attention to the relationship between the first displayed content and the associated entity, and the associated entity includes an entity with a relationship in the knowledge graph of the first displayed content. The contextual intention represents the user's attention to the context of the first operation.

[0268] The representation determination module 320 is used to obtain embedded representation information based on the user's relationship intention and context intention through a graph neural network.

[0269] The recommendation module 330 is configured to output the second display content through a recommendation model at least according to the embedded representation information.

[0270] In one possible implementation, the representation determination module 320 is specifically used to: input the graph representation into a graph neural network; the graph representation includes user nodes, item nodes, fusion intent nodes, and connection edges; the user node represents the user, the item node represents the first display content, and the fusion intent node represents the user's fusion intent; the fusion intent represents the result of the fusion of relationship intent and context intent; the connection edge is used to connect the user node and the fusion intent node, the fusion intent node and the item node; the graph neural network is used to update the features of the nodes represented by the graph, and output embedded representation information; the embedded representation information includes the user representation output by the user node and the item representation output by the item node.

[0271] In another possible implementation, the characterization determination module 320 is specifically used to: update the features of the user node based on the features of the item node and the fusion intention node; and update the features of the item node based on the features of the user node and the fusion intention node.

[0272] In another possible implementation, the representation determination module 320 is specifically used for: the graph representation also includes context intent nodes and context nodes; the connecting edges are also used to connect user nodes and context intent nodes, context intent nodes and context nodes, and item nodes and context intent nodes; the context intent nodes represent the user's context intent, and the context nodes represent the context of the first operation; the embedded representation information also includes the context representation of the context node output.

[0273] In another possible implementation, the characterization determination module 320 is specifically used to: update the features of the user node based on the features of the item node and the fused intent node; update the features of the item node based on the features of the user node and the fused intent node; after updating the features of the context intent node based on the features of the user node and the item node, update the features of the context node based on the context intent and the features of the context intent node; and / or update the features of the context intent node based on the features after the features of the context node are updated, so that the features of the fused intent node are updated based on the features of the context intent node.

[0274] In another possible implementation, the graph representation also includes entity nodes and relationship edges; entity nodes represent associated entities; relationship edges represent the relationship between items and associated entities; the initial features of the entity nodes are the information of the associated entities in the knowledge graph; relationship edges are used to connect item nodes and entity nodes; the representation determination module 320 is specifically used to: update the features of the entity nodes based on the features of the item nodes and relationship edges; update the features of the item nodes based on the features of the entity nodes and relationship edges.

[0275] In another possible implementation, the intent determination module 310 is specifically used to obtain a relationship intent based on the user's attention to one or more relationship sub-intents; the relationship sub-intent represents the degree of attention to the associated entity.

[0276] In another possible implementation, the contextual intent further represents the user's attention to the context of the fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

[0277] Fig.25 A schematic diagram of the structure of a computing device provided in an embodiment of the present application.

[0278] like Fig.25 As shown, the computing device 400 includes at least one processor 410, a memory 420 and a communication interface 430. The processor 410, the memory 420 and the communication interface 430 are connected in communication, and the communication connection can be achieved by wired means (such as a bus) or by wireless means. The communication interface 430 is used to receive data sent by other devices (such as packet information sent by a cloud management platform); the memory 420 stores computer instructions, and the processor 410 executes the computer instructions to execute the method in the aforementioned method embodiment.

[0279] It should be understood that in the embodiment of the present application, the processor 410 may be a central processing unit CPU, and the processor 1401 may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0280] The memory 420 may include a read-only memory and a random access memory, and provides instructions and data to the processor 410. The memory 420 may also include a nonvolatile random access memory.

[0281] The memory 420 may be a volatile memory or a nonvolatile memory, or may include both volatile and nonvolatile memories. Among them, the nonvolatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct ram bus RAM (DR RAM).

[0282] It should be understood that the computing device 400 according to the embodiment of the present application can execute the method mentioned in the embodiment of the present application. The detailed description of the implementation of the method is shown above, and for the sake of brevity, it will not be repeated here.

[0283] An embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer instructions are executed by a processor, the above-mentioned technical solution is implemented.

[0284] An embodiment of the present application provides a chip, which includes at least one processor and an interface, wherein the at least one processor determines program instructions or data through the interface; the at least one processor is used to execute the program instructions to implement the technical solution mentioned above.

[0285] An embodiment of the present application provides a computer program or a computer program product, which includes instructions. When the instructions are executed, the computer executes the technical solution mentioned above.

[0286] Those of ordinary skill in the art should further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0287] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented by hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0288] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only the specific implementation method of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A display method, characterized in that: include: The first device displays a first user interface, wherein the first user interface includes: first display content; In response to the first device detecting a first operation acting on the first display content, a second user interface is displayed; the second user interface includes: a second display content; the second display content is related to a relational intent and a contextual intent; wherein the second display content is a result of a recommendation model making a recommendation based at least on embedded representation information; the embedded representation information is representation information output by a graph neural network based on the relational intent and the contextual intent; the relational intent represents the user's attention to the relationship between the first display content and associated entities, and the associated entities include entities with relationships in the knowledge graph of the first display content; the contextual intent represents the user's attention to the context of the first operation.

2. The display method according to claim 1, characterized in that: After the first device displays the second user interface, the display method further includes: After the first device detects the second operation, it displays a third user interface; the third user interface displays: at least one of the associated entities and the user's attention level to the associated entity; and / or at least one of the contexts of the first operation and the user's attention level to the context of the first operation.

3. The display method according to claim 1 or 2, characterized in that: After the first device displays the second user interface, the display method further includes: After detecting the third operation, the first device displays a fourth user interface; the fourth user interface displays: at least one of the first display contents, and / or at least one of the associated entities, and / or at least one context of the first operation.

4. The display method according to any one of claims 1 to 3, characterized in that: The embedded representation information is the representation information output by the graph neural network based on the relationship intent and the context intent, including: The graph representation of the graph neural network input includes a user node, an item node, a fusion intention node, and a connection edge; the user node represents the user, the item node represents the first display content, and the fusion intention node represents the fusion intention of the user; the fusion intention represents the result of the fusion of the relationship intention and the context intention; the connection edge is used to connect the user node and the fusion intention node, and the fusion intention node and the item node; The graph neural network is used to update the features of the nodes represented by the graph and output the embedded representation information; the embedded representation information includes the user representation output by the user node and the item representation output by the item node.

5. The display method according to claim 4, characterized in that: The updating of features of the nodes represented by the graph includes: Based on the features of the item node and the fusion intention node, updating the features of the user node; Based on the features of the user node and the fusion intent node, the features of the item node are updated.

6. The display method according to claim 4, characterized in that: The graph representation of the graph neural network input also includes a contextual intent node and a context node; the connecting edge is also used to connect the user node and the contextual intent node, the contextual intent node and the context node, and the item node and the contextual intent node; The context intention node represents the context intention, and the context node represents the context of the first operation; The embedded representation information also includes a context representation output by the context node.

7. The display method according to claim 6, characterized in that: The updating of features of the nodes represented by the graph includes: Based on the features of the item node and the fusion intention node, updating the features of the user node; Based on the features of the user node and the fusion intent node, the features of the item node are updated; after the features of the context intent node are updated based on the information of the user node and the item node, the features of the context node are updated based on the context intent and the features of the context intent node; And / or based on the updated features of the context node, update the features of the context intent node, so that based on the features of the context intent node, update the features of the fusion intent node.

8. The display method according to any one of claims 4 to 7, characterized in that: The graph representation further includes entity nodes and relationship connection edges; the relationship connection edges are used to connect the item nodes and the entity nodes; the entity nodes represent the associated entities; the relationship connection edges represent the relationship between the item and the associated entity; The initial feature of the entity node is the information of the associated entity in the knowledge graph; The updating of features of the nodes represented by the graph further includes: Based on the features of the item node and the relationship connection edge, updating the features of the entity node; Based on the features of the entity node and the relationship connection edge, the features of the item node are updated.

9. The display method according to any one of claims 1 to 8, characterized in that: The relationship intent represents the user's attention to the relationship between the first display content and the associated entity, including: obtaining the relationship intent based on the user's attention to one or more relationship sub-intentions; the relationship sub-intention represents the attention to the associated entity.

10. The system according to any one of claims 1 to 9, characterized in that: The contextual intent also represents the user's attention to the context of a fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

11. A recommendation method, characterized in that: include: Determine the relationship intention and context intention of the user based on the user's historical interaction data; The historical interaction data includes the first operation of the user on the first displayed content and the context of the first operation; the relationship intent represents the user's attention to the relationship between the first displayed content and the associated entity, and the associated entity includes an entity with a relationship in the knowledge graph of the first displayed content; the contextual intent represents the user's attention to the context of the first operation; Based on the relational intention and contextual intention of the user, embedded representation information is obtained through a graph neural network; The recommendation model outputs second display content based at least on the embedded representation information.

12. The recommendation method according to claim 11, characterized in that: The method of obtaining embedded representation information based on the relational intention and contextual intention of the user through a graph neural network includes: inputting a graph representation into the graph neural network; the graph representation includes a user node, an item node, a fusion intention node, and a connection edge; the user node represents the user, the item node represents the first display content, and the fusion intention node represents the fusion intention of the user; the fusion intention represents the result of the fusion of the relational intention and the contextual intention; the connection edge is used to connect the user node and the fusion intention node, and the fusion intention node and the item node; The graph neural network updates the features of the nodes represented by the graph and outputs the embedded representation information; the embedded representation information includes the user representation output by the user node and the item representation output by the item node.

13. The recommendation method according to claim 12, characterized in that: The graph neural network updates features of nodes represented by the graph, including: Based on the features of the item node and the fusion intention node, updating the features of the user node; Based on the features of the user node and the fusion intent node, the features of the item node are updated.

14. The recommendation method according to claim 12, characterized in that: The embedding representation information is obtained based on the relational intention and contextual intention of the user through the graph neural network, including: The graph representation also includes a contextual intent node and a context node; the connecting edge is also used to connect the user node and the contextual intent node, the contextual intent node and the context node, the item node and the contextual intent node; the contextual intent node represents the contextual intent of the user, and the context node represents the context of the first operation; The embedded representation information also includes a context representation output by the context node.

15. The recommendation method according to claim 14, characterized in that: The graph neural network updates features of nodes represented by the graph, including: Based on the features of the item node and the fusion intention node, updating the features of the user node; Based on the features of the user node and the fusion intention node, updating the features of the item node; After updating the features of the contextual intent node based on the features of the user node and the item node, updating the features of the contextual node based on the contextual intent and the features of the contextual intent node; And / or based on the updated features of the context node, update the features of the context intent node, so that based on the features of the context intent node, update the features of the fusion intent node.

16. The recommendation method according to claims 12-15, characterized in that: The graph representation further includes entity nodes and relationship connection edges; the entity nodes represent the associated entities; the relationship connection edges represent the relationship between the items and the associated entities; the initial features of the entity nodes are the information of the associated entities in the knowledge graph; The relationship connection edge is used to connect the item node and the entity node; The graph neural network updates the features of the nodes represented by the graph, and further includes: Based on the features of the item node and the relationship connection edge, updating the features of the entity node; Based on the features of the entity node and the relationship connection edge, the features of the item node are updated.

17. The recommendation method according to any one of claims 11 to 16, characterized in that: The relationship intent represents the user's attention to the relationship between the item and the associated entity, including: The relationship intent is obtained based on the user's attention degree to one or more relationship sub-intents; the relationship sub-intent represents the attention degree to the related entity.

18. The recommendation method according to any one of claims 11 to 17, characterized in that: The contextual intent also represents the user's attention to the context of a fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

19. A recommendation device, characterized in that: include: An intention determination module, used to determine the relationship intention and context intention of the user based on the user's historical interaction data; The historical interaction data includes the first operation of the user on the first displayed content and the context of the first operation; the relationship intent represents the user's attention to the relationship between the first displayed content and the associated entity, and the associated entity includes an entity with a relationship in the knowledge graph of the first displayed content; the contextual intent represents the user's attention to the context of the first operation; A representation determination module, used to obtain embedded representation information through a graph neural network based on the relationship intention and context intention of the user; A recommendation module is used to output second display content through a recommendation model at least according to the embedded representation information.

20. The recommendation device according to claim 19, characterized in that The characterization determination module is specifically used to: Inputting a graph representation into a graph neural network; the graph representation includes a user node, an item node, a fusion intention node, and a connection edge; the user node represents the user, the item node represents the first display content, and the fusion intention node represents the fusion intention of the user; The fusion intent represents the result of the fusion of the relationship intent and the context intent; The connection edge is used to connect the user node and the fusion intention node, and the fusion intention node and the item node; Performing feature updates on the nodes represented by the graph through the graph neural network, and outputting the embedded representation information; The embedded representation information includes the user representation output by the user node and the item representation output by the item node.

21. The recommendation device according to claim 20, characterized in that: The characterization determination module is specifically used to: Based on the features of the item node and the fusion intention node, updating the features of the user node; Based on the features of the user node and the fusion intent node, the features of the item node are updated.

22. The recommendation device according to claim 20, characterized in that The characterization determination module is specifically used to: The graph representation also includes a contextual intent node and a context node; the connecting edge is also used to connect the user node and the contextual intent node, the contextual intent node and the context node, and the item node and the contextual intent node; The contextual intention node represents the contextual intention of the user, and the context node represents the context of the first operation; The embedded representation information also includes a context representation output by the context node.

23. The recommendation device according to claim 22, characterized in that: The characterization determination module is specifically used to: Based on the features of the item node and the fusion intention node, updating the features of the user node; Based on the features of the user node and the fusion intention node, updating the features of the item node; After updating the features of the contextual intent node based on the features of the user node and the item node, updating the features of the contextual node based on the contextual intent and the features of the contextual intent node; And / or based on the updated features of the context node, update the features of the context intent node, so that based on the features of the context intent node, update the features of the fusion intent node.

24. The recommendation device according to claims 20-23, characterized in that The graph representation also includes entity nodes and relationship connection edges; the entity nodes represent the associated entities; the relationship connection edges represent the relationship between the items and the associated entities; the initial feature of the entity nodes is the information of the associated entities in the knowledge graph; the relationship connection edges are used to connect the item nodes and the entity nodes; The characterization determination module is specifically used to: Based on the features of the item node and the relationship connection edge, updating the features of the entity node; Based on the features of the entity node and the relationship connection edge, the features of the item node are updated.

25. The recommendation device according to any one of claims 19 to 24, characterized in that: The intention determination module is specifically used to: The relationship intent is obtained based on the user's attention degree to one or more relationship sub-intents; the relationship sub-intent represents the attention degree to the related entity.

26. The recommendation device according to any one of claims 19 to 25, characterized in that: The contextual intent also represents the user's attention to the context of a fourth operation; the fourth operation is an operation performed by the user on the first displayed content before the first operation.

27. A computing device, characterized in that It comprises a memory and a processor, wherein the memory stores instructions, and when the instructions are executed by the processor, the display method as described in any one of claims 1 to 10 is implemented, and / or the recommendation method as described in any one of claims 11 to 18 is implemented, and / or the recommendation device as described in any one of claims 19 to 26 is implemented.

28. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the display method described in any one of claims 1 to 10 is implemented, and / or the recommendation method described in any one of claims 11 to 18 is implemented, and / or the recommendation device described in any one of claims 19 to 26 is implemented.

29. A computer program product comprising program instructions, characterized in that: When the program instructions are executed by a computer, the computer executes the display method as described in any one of claims 1 to 10, and / or executes the recommendation method as described in any one of claims 11 to 18, and / or implements the recommendation device as described in any one of claims 19 to 26.

Images