Resource recall methods, devices, servers, and storage media

Abstract

This disclosure relates to a resource retrieval method, apparatus, server, and storage medium. The method includes: constructing multiple heterogeneous graphs based on a historical interaction dataset between multiple users and multiple resources; the multiple heterogeneous graphs include a behavioral graph representing the interaction records between users and resources, a modal graph representing the relationships between various resources under different modal features, and a knowledge graph representing the relationships between various resources under preset attribute information; determining the target embedding representation of each user and the target embedding representation of each resource based on the connection relationships between nodes in the heterogeneous graphs and the initial embedding representation of each node; for each resource, calculating a retrieval score for the resource based on the target embedding representation of the resource and the target embedding representation of each user; and determining the retrieved resource from the multiple resources based on the retrieval score of each resource. This method can maintain high-quality retrieval capabilities even in data sparse and cold-start environments.

Description

Technical Field

[0001] This disclosure relates to the field of recommendation technology, and in particular to a resource retrieval method, apparatus, server, storage medium, and program product. Background Technology

[0002] In the recommendation field, there are two main methods for resource retrieval in recommendation systems. One method is to apply graph neural networks for retrieval modeling. Specifically, this involves constructing a user-item graph and using simplified graph convolution operations to learn node representations, thereby achieving collaborative filtering recommendations between users and resources. The other method is multimodal recommendation, which initializes each resource node as a representation that integrates multimodal features such as images, text, and tags. Then, it performs graph neural network propagation operations based on the user-item graph. Modal information does not participate in the edge construction or path control of the graph structure itself.

[0003] However, both of these methods rely solely on user-resource interactions during propagation. For new users with no behavioral records or sparse behavioral data, the lack of effective interaction with a large amount of long-tail content can lead to graph structure breaks, making it impossible to establish a propagation path and impacting recall capabilities. Summary of the Invention

[0004] This disclosure provides a resource retrieval method, apparatus, server, storage medium, and program product to at least solve the problem of poor retrieval capability for new users with no behavioral records or in situations with sparse behavioral data in related technologies. The technical solution of this disclosure is as follows:

[0005] According to a first aspect of the present disclosure, a resource recall method is provided, comprising:

[0006] Based on the historical interaction dataset between multiple users and multiple resources, multiple heterogeneous graphs are constructed. The multiple heterogeneous graphs include a behavioral graph representing the interaction records between users and resources, a modal graph representing the relationship between each resource under different modal features, and a knowledge graph representing the relationship between each resource under preset attribute information.

[0007] Based on the connection relationships between nodes in each heterogeneous graph and the initial embedding representation of each node, the target embedding representation of each user and the target embedding representation of each resource are determined.

[0008] For each resource, a recall score for that resource is calculated based on the target embedding representation of the resource and the target embedding representation of each user.

[0009] Based on the recall score of each resource, the recalled resources are determined from the plurality of resources.

[0010] In an exemplary embodiment, each modal feature corresponds to a modality graph; the process of constructing the modality graph based on the historical interaction dataset includes:

[0011] For each modal feature, the similarity between every two resources among the multiple resources under that modal feature is obtained;

[0012] Using each resource as a node, establish connection edges between nodes corresponding to two resources with similarity higher than a threshold to obtain the modality graph under this modality feature.

[0013] In one exemplary embodiment, there are multiple preset attribute information; the process of constructing the knowledge graph based on the historical interaction dataset includes:

[0014] Determine the relationship between each resource and each attribute information, as well as the relationship between each attribute information;

[0015] Using each resource and each attribute information as nodes, connection edges are established between related resources and attribute information, and between related attribute information, to obtain the knowledge graph.

[0016] In an exemplary embodiment, determining the target embedding representation for each user and the target embedding representation for each resource based on the connection relationships between nodes in the heterogeneous graphs and the initial embedding representations of each node includes:

[0017] By using a preset propagation control function, propagation weights are assigned between two nodes that are connected in each heterogeneous graph; wherein, the gated network corresponding to each heterogeneous graph is different in the propagation control function.

[0018] Based on the propagation weights and the initial embedding representations of each node, a multi-layer joint propagation operation of a graph neural network is performed on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource.

[0019] In an exemplary embodiment, the step of performing a multi-layer joint propagation operation of a graph neural network on each heterogeneous graph based on the propagation weights and the initial embedding representations of each node to obtain the target embedding representation of each user and the target embedding representation of each resource includes:

[0020] For any one of the multiple users and multiple resources, in each layer, the object embedding representation of that object in each heterogeneous graph is determined;

[0021] The object embedding representations of each heterogeneous graph are fused to obtain the single-layer embedding representation of the object in each layer.

[0022] The target embedding representation of the object is obtained by weighted averaging of the single-layer embedding representations obtained from each layer.

[0023] In an exemplary embodiment, assigning propagation weights between two nodes connected in each heterogeneous graph using a preset propagation control function includes:

[0024] For modal graphs and knowledge graphs in various heterogeneous graphs, propagation weights are assigned between two nodes with a connection relationship through a preset propagation control function, in a way that the propagation weights are positively correlated with semantic relevance.

[0025] In one exemplary embodiment, the method further includes:

[0026] For users who have not interacted with any resources among the multiple users, or resources that have not been interacted with by any user among the multiple resources, the target embedding representation is determined based on their modal graph and knowledge graph.

[0027] According to a second aspect of the present disclosure, a resource recall apparatus is provided, comprising:

[0028] The heterogeneous graph construction unit is configured to construct multiple heterogeneous graphs based on a historical interaction dataset between multiple users and multiple resources. The multiple heterogeneous graphs include a behavioral graph representing the interaction records between users and resources, a modal graph representing the relationship between each resource under different modal features, and a knowledge graph representing the relationship between each resource under preset attribute information.

[0029] The information propagation unit is configured to perform the determination of the target embedding representation for each user and the target embedding representation for each resource based on the connection relationships between nodes in each heterogeneous graph and the initial embedding representation of each node.

[0030] The scoring unit is configured to perform a recall score for each resource based on the target embedding representation of the resource and the target embedding representation of each user.

[0031] The resource filtering unit is configured to perform a recall score based on each resource to determine the recall resources from the plurality of resources.

[0032] In an exemplary embodiment, each modal feature corresponds to a modal graph; the heterogeneous graph construction unit is further configured to perform the following: for each modal feature, obtain the similarity between every two resources among the plurality of resources under the modal feature; and establish connection edges between the nodes corresponding to two resources with similarity higher than a threshold, using each resource as a node, to obtain the modal graph under the modal feature.

[0033] In an exemplary embodiment, there are multiple preset attribute information; the heterogeneous graph construction unit is further configured to perform the following: determine the correlation between each resource and each attribute information, and the correlation between each attribute information; using each resource and each attribute information as nodes, establish connection edges between related resources and attribute information, and establish connection edges between related attribute information, to obtain the knowledge graph.

[0034] In an exemplary embodiment, the information propagation unit is further configured to execute a preset propagation control function to assign propagation weights between two nodes that are connected in each heterogeneous graph; wherein, in the propagation control function, the gating networks corresponding to each heterogeneous graph are different; based on the propagation weights and the initial embedding representations of each node, a multi-layer joint propagation operation of a graph neural network is performed on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource.

[0035] In an exemplary embodiment, the information dissemination unit is further configured to perform the following actions for any one of the plurality of users and the plurality of resources: in each layer, determining the object embedding representation of the object in each heterogeneous graph; fusing the object embedding representations in the various heterogeneous graphs to obtain a single-layer embedding representation of the object in each layer; and performing a weighted average of the single-layer embedding representations obtained in each layer to obtain a target embedding representation of the object.

[0036] In an exemplary embodiment, the information propagation unit is further configured to perform propagation weights on modal graphs and knowledge graphs in each heterogeneous graph, using a preset propagation control function, in a manner that the propagation weights are positively correlated with semantic relevance, and assign propagation weights to two nodes that have a connection relationship.

[0037] In one exemplary embodiment, the information dissemination unit is further configured to determine a target embedding representation based on its modal graph and knowledge graph for users among the plurality of users who have not interacted with any resource, or for resources among the plurality of resources that have not been interacted with by any user.

[0038] According to a third aspect of the present disclosure, a server is provided, comprising:

[0039] processor;

[0040] Memory used to store the processor's executable instructions;

[0041] The processor is configured to execute the instructions to implement the method as described in any of the preceding methods.

[0042] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of a server, enables the server to perform the method described in any of the preceding claims.

[0043] According to a fifth aspect of the present disclosure, a computer program product is provided, the computer program product including instructions that, when executed by a processor of a server, enable the server to perform the method as described in any of the preceding claims.

[0044] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0045] This disclosure constructs multiple heterogeneous graphs based on historical interaction datasets between multiple users and multiple resources. These graphs include a behavior graph representing user-resource interaction records, a modality graph representing the relationships between resources under different modal features, and a knowledge graph representing the relationships between resources under preset attribute information. The target embedding representations for each user and each resource are jointly determined based on these heterogeneous graphs. Furthermore, based on the target embedding representations of resources and each user, a recall score for the resource is calculated, thereby identifying the recalled resource from multiple resources. This multi-source information joint recall method, combining behavior graphs, modality graphs, and knowledge graph structures, no longer relies on a single relationship to connect users and resources. Instead, it uses three types of edges—user behavior, modality similarity, and co-occurrence of knowledge entities—to construct heterogeneous graphs. This enables richer and more semantically consistent representation learning for nodes, maintaining high-quality recall capabilities even in data sparsity and cold-start environments. It significantly improves recommendation accuracy and system stability, solving the problems of behavior sparsity, resource cold start, and semantic disconnect in complex recommendation scenarios.

[0046] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0048] Figure 1 This is a flowchart illustrating a resource retrieval method according to an exemplary embodiment.

[0049] Figure 2 This is a schematic diagram illustrating a behavior diagram according to an exemplary embodiment.

[0050] Figure 3 This is a schematic diagram of a modal diagram according to an exemplary embodiment.

[0051] Figure 4 This is a schematic diagram of a knowledge graph according to an exemplary embodiment.

[0052] Figure 5 This is a schematic diagram of a knowledge graph according to another exemplary embodiment.

[0053] Figure 6 This is a schematic diagram illustrating a process for determining a target embedding representation based on a graph neural network, according to an exemplary embodiment.

[0054] Figure 7 This is a flowchart illustrating a resource retrieval method according to another exemplary embodiment.

[0055] Figure 8 This is a structural block diagram of a resource recall device according to an exemplary embodiment.

[0056] Figure 9 This is a block diagram illustrating a server according to an exemplary embodiment. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0058] It should be noted that the embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims. It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0059] It is understandable that recommendation systems in related technologies often rely on user-resource interaction to construct graph structures and use graph neural networks to generate embedded nodes. However, such methods face three difficulties in actual platforms: (1) behavioral sparsity: a large amount of long-tail content lacks effective interaction with new users, resulting in graph structure breaks and the inability to establish propagation paths; (2) modal heterogeneity: content usually has multimodal attributes such as images, text, and audio, and a single structure cannot carry all the information; (3) semantic disconnect: there are a large amount of semantically related but structurally unconnected content, such as "same type of works" or "same author", which are difficult to model in the graph.

[0060] Based on this, this application proposes a three-source fusion graph recall method. The core idea is to construct a triple graph structure (behavioral graph + modal graph + knowledge graph) and perform joint propagation control during the graph neural network propagation process to improve the diversity and semantic consistency of information paths. In other words, this method no longer relies on a single relationship to connect users and resources, but uses three types of edges—user behavior, modal similarity, and co-occurrence of knowledge entities—to construct a super graph, and controls their participation in the graph neural network through a designed propagation mechanism, thereby achieving structural enhancement and semantic completion.

[0061] In one exemplary embodiment, such as Figure 1 As shown, a resource retrieval method is provided. This embodiment illustrates the method by applying it to a server, which can be implemented using a standalone server or a server cluster composed of multiple servers. In this embodiment, the method includes the following steps:

[0062] In step S110, multiple heterogeneous graphs are constructed based on the historical interaction dataset between multiple users and multiple resources. The multiple heterogeneous graphs include a behavior graph representing the interaction records between users and resources, a modality graph representing the relationship between each resource under different modal features, and a knowledge graph representing the relationship between each resource under preset attribute information.

[0063] Here, a resource represents a content entity that can be recommended to a user; for example, a resource can be a video.

[0064] Behavioral graphs are formed by establishing interaction edges between users and resources based on user behaviors such as clicks, likes, and views. Behavioral graphs have two types of nodes: user nodes and resource nodes, connected by edges. It is established between user nodes and resource nodes, and only by the user node. With resources This is built upon prior interaction (such as clicks or purchases). Assume the following interaction data: User: ,resource: Interaction log: Clicked ; Clicked The constructed behavior graph is as follows Figure 2 As shown, in the user With resources A connection edge is established between them, in the user With resources There are connecting edges between them.

[0065] Modality graphs establish modal similarity connections between resources based on their modal features (such as images and text). The nodes in the modality graph are resource nodes. A modality graph is generated for each modality; for example, a modality graph can be generated for text features, and a modality graph can be generated for image features. Figure 3 As shown, Figure 3 Figures (a) and (b) in the figure represent modal diagrams under text features and image features, respectively. Figure 3 Figure (a) in the diagram represents the resources. With resources ,resource ,resource The text features are similar, resources With resources ,resource (a) Textual features are similar; (b) Graph representation resources With resources ,resource The images have similar features, and the resources With resources Their image features are similar.

[0066] In this context, a knowledge graph represents a semantic network constructed based on resource attribute information (such as tags, authors, categories, styles, etc.). Nodes in a knowledge graph include resource nodes and entity nodes corresponding to each attribute. Resources are indirectly connected through sharing attribute information, and entities can further establish weak connections or aggregations through co-occurrence relationships. For example, ... Figure 4 As shown, resources ,resource and resources Indirect connections can be established between them through shared label Y, and there is an entity connection between category Y and author B.

[0067] It is understandable that traditional recall methods only utilize user-resource interaction edges to construct graph structures, failing to incorporate modal information such as images and text. This results in poor recall performance on cold-start resources or users with no interaction behavior. Therefore, in constructing heterogeneous graphs, this step, in addition to constructing behavioral graphs, also introduces modal graphs and knowledge graphs to jointly represent the relationship between users and resources, breaking the limitation of traditional recommendation systems that rely solely on user-resource behavior to construct graph structures.

[0068] In step S120, based on the connection relationships between nodes in each heterogeneous graph and the initial embedding representation of each node, the target embedding representation of each user and the target embedding representation of each resource are determined.

[0069] User embeddings and item embeddings are numerical representations that map users and resources to a low-dimensional vector space. These embeddings can capture the latent semantics of user preferences and resource features.

[0070] In the specific implementation, the target embedding representations (or final embedding representations) of users and resources are determined through the information propagation mechanism of graph neural networks. Specifically, the information propagation operation of graph neural networks mainly includes four steps: initializing node embeddings, neighbor aggregation, multi-layer propagation, and final embedding generation. Initializing node embeddings involves initializing the embedding of each node in each heterogeneous graph. For example, the initial embedding representation of a user could be their interests, age, occupation, etc., while the embedding representation of a resource could be its features (such as content, description, etc.). Neighbor aggregation occurs in each layer of the graph neural network, where nodes update their own representations by aggregating information from their neighbors. Multi-layer propagation combines the user and resource embedding representations from each layer with the embedding representations obtained from the previous layer. Finally, the user and resource embedding representations from all layers are weighted and averaged to obtain the target embedding representation for each user and each resource.

[0071] Through multi-layer propagation, user embedding can capture resources liked by similar users, similar resources of the liked resources, and resources with similar attribute information to the liked resources. Even if some users and resources do not have direct interaction, indirect connections can be established through graph structure to alleviate the sparsity problem of interaction data. At the same time, dynamic interest diffusion can be realized, allowing user interests to spread from directly interacting resources to related resources (such as u1→i2→u2→i3, etc.).

[0072] In step S130, for each resource, a recall score for the resource is calculated based on the target embedding representation of the resource and the target embedding representation of each user.

[0073] The recall score can be used to measure the matching degree between users and resources, thereby filtering out recall resources with a high matching degree from a large number of candidate resources.

[0074] For example, for each resource, the vector dot product of each user's target embedding representation and the resource's target embedding representation can be calculated separately to obtain multiple scores for that resource. The average of these scores is then calculated to obtain the recall score for that resource. This process can be expressed by the following formula:

[0075] (1)

[0076] in, Indicates user Resources The recall score, Indicates user The target embedding representation, Representing resources The target embedding representation.

[0077] For example, given m resources and n users, for any resource, calculate the vector inner product of the resource and the target embedding representation of each user to obtain n scores for that resource. Then, calculate the average of these n scores to obtain the recall score for that resource.

[0078] In some embodiments, to improve efficiency, a predetermined proportion of users and resources can be selected from multiple users to calculate the recall score, without requiring the target embedding representation of each user to participate in the calculation of the recall score for each resource.

[0079] In step S140, based on the recall score of each resource, the recalled resources are determined from multiple resources.

[0080] In practice, a higher recall score indicates a greater probability of user interest in the resource and a higher recommendation rate. Therefore, the top N resources with the highest recall scores can be selected from multiple resources and used as the recall resources, arranged in descending order of recall score.

[0081] In the aforementioned resource retrieval method, based on historical interaction datasets between multiple users and multiple resources, multiple heterogeneous graphs are constructed, including a behavior graph representing user-resource interaction records, a modality graph representing the relationships of various resources under different modal features, and a knowledge graph representing the relationships of various resources under preset attribute information. These heterogeneous graphs are used to jointly determine the target embedding representation for each user and each resource. Furthermore, based on the target embedding representations of resources and each user, a retrieval score for the resource is calculated, thereby identifying the retrieved resource from multiple resources. This multi-source information joint retrieval method, combining behavior graphs, modality graphs, and knowledge graph structures, no longer relies on a single relationship to connect users and resources. Instead, it uses three types of edges—user behavior, modality similarity, and co-occurrence of knowledge entities—to construct heterogeneous graphs. This enables richer and more semantically consistent representation learning for nodes, maintaining high-quality retrieval capabilities even in data sparsity and cold-start environments. It significantly improves recommendation accuracy and system stability, solving the problems of behavior sparsity, resource cold start, and semantic disconnect in complex recommendation scenarios.

[0082] It is understood that modality graphs connect different resources through modalities. The connection relationships between resources are different under different modal features. Therefore, it is necessary to construct a corresponding modality graph for each modality feature to represent the connection relationships of each resource under different modality features. In an exemplary embodiment, in step S110, the process of constructing a modality graph based on a historical interaction dataset includes: for each modality feature, obtaining the similarity between every two resources under that modality feature; using each resource as a node, establishing connection edges between nodes corresponding to two resources with similarity higher than a threshold to obtain the modality graph under that modality feature.

[0083] Specifically, we identify the various modal features for which we need to construct the modal graph, such as text features and image features. For each modal feature, we calculate the similarity between any two resources under that modal feature. We establish connecting edges between nodes corresponding to two resources with similarity higher than a threshold, thereby forming the modal graph under that modal feature.

[0084] For example, suppose resources With resources ,resource ,resource If the similarity of the text features is higher than the threshold, then it can be used in the resources. With resources ,resource ,resource Establish connection edges between them, resources With resources ,resource If the similarity of the text features is higher than the threshold, then it can be used in the resources. With resources ,resource Connections are also established between them, resulting in the modal diagram as follows: Figure 3 As shown in Figure (a). For example, suppose resources... With resources ,resource If the similarity of the image features is higher than the threshold, then it can be used in the resource With resources ,resource Establish connecting edges between them; resources With resources If the similarity of the image features is higher than the threshold, then it can be used in the resource With resources Establish connecting edges between them.

[0085] In this embodiment, by using modal graphs of various resources under multiple modal features, the association between resources can be established based on the multimodal features of resources. This allows subsequent path selection to be guided based on modal features without relying solely on user and resource behavior interactions, thereby improving the representation capabilities of users and resources and achieving high-quality resource retrieval.

[0086] It is understood that a knowledge graph is constructed based on resource nodes and their associated attribute information. Attribute information can be various entities such as tags, authors, and styles; that is, a resource can have multiple attribute information. In an exemplary embodiment, the process of constructing a knowledge graph based on a historical interaction dataset includes: determining the association between each resource and each attribute information, as well as the association between each attribute information; using each resource and each attribute information as nodes, establishing connection edges between related resources and attribute information, and establishing connection edges between related attribute information, to obtain the knowledge graph.

[0087] Specifically, when constructing a knowledge graph, each resource and each attribute information is used as a node. After determining the relationship between each resource and each attribute information, connection edges are established between each resource and each attribute information. If there are also relationships between the attribute information, then connection edges are established between the related attribute information.

[0088] For example, if: -Author A-Style C Author A - Style B -Author B-Style B, then in Establish connection edges between author A and style C, in Establish connection edges between author A and style B, in Establish connection edges between author B and style B, such as Figure 5 As shown. If there is a relationship between author A and style C, then a connection edge is also established between author A and style C, thereby obtaining a knowledge graph representing the relationship between various resources under preset attribute information.

[0089] In some embodiments, considering the varying degrees of correlation between different attribute information, the connections between attribute information can be weak or strong. For example, if author A's main style is style B, and some of his works are in style C, then author A has a strong connection with style B, and author A has a weak connection with style C.

[0090] In this embodiment, by constructing a knowledge graph of each resource based on its attribute information, semantic association modeling can be achieved: for example, establishing connection edges between semantically similar resources such as "same author" or "same style". This method, by modeling the semantic structural relationships between complex resources, can improve the richness and generalization ability of resource embedding representation, thereby improving the quality of resource retrieval.

[0091] In one exemplary embodiment, such as Figure 6 As shown, step S120, based on the connection relationships between nodes in each heterogeneous graph and the initial embedding representation of each node, determines the target embedding representation for each user and the target embedding representation for each resource, including:

[0092] Step S121: Assign propagation weights between two nodes that are connected in each heterogeneous graph using a preset propagation control function;

[0093] Step S122: Based on the propagation weights and the initial embedding representations of each node, perform multi-layer joint propagation operations of graph neural networks on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource.

[0094] In the propagation control function, the gated networks corresponding to each heterogeneous graph are different. Refer to the following formula (2) for an example of the propagation control function:

[0095] (2)

[0096] in, and This represents the embedding of two nodes. For different path types of heterogeneous graphs In a gated network, the core task of an MLP is to map input features to gate weights. It can be understood that the propagation weights (or attention) of the edges connecting two nodes determine the importance allocation of neighboring nodes during information propagation. In this embodiment, different heterogeneous graphs correspond to different gated networks, assigning different propagation weights to the edges connecting these heterogeneous graphs to control the propagation weights of different heterogeneous graphs and ensure the accuracy and effectiveness of the propagation process.

[0097] In some embodiments, a Graph Attention Network (GAT) model can be used to calculate attention scores for the three types of connection edges in the three types of heterogeneous graphs, which are then used as propagation weights to dynamically adjust the propagation path. It should be noted that the graph attention model learns attention for each edge individually and does not distinguish between edge types.

[0098] Specifically, the propagation operation of a graph neural network involves the graph neural network transmitting and aggregating information based on the interaction between users and resources, the relationships between resource modalities, and the relationships between resources and attribute information, thereby learning richer user and resource representations. The core of this process is using the graph structure to capture collaborative filtering signals (such as "similar users like similar resources"). The propagation hierarchy of a graph neural network is generally 1-3 layers, with each layer including three key operations: message passing, aggregation, and update. Message passing refers to each node generating and sending "messages" to its neighboring nodes; these messages are typically determined by the node's own features, the features of its neighbors, and the features of the edges (if any). Aggregation refers to each node combining messages from its neighbors into a comprehensive neighborhood information. Update refers to combining the node's previous features with the aggregated neighborhood information to generate new features for the node.

[0099] More specifically, in the first layer, nodes only aggregate information from their direct neighbors; in the second layer, nodes aggregate information from the neighbors of their neighbors; and so on, so that user nodes and resource nodes can capture the structure and features of their neighboring nodes, thus achieving hierarchical feature learning.

[0100] For example, suppose in the first layer, node A aggregates information from nodes B and C, and node B aggregates information from node D. Then in the second layer, node A indirectly aggregates information from node D.

[0101] In this embodiment, a propagation control function is designed to create different gating networks for different types of heterogeneous graphs, thereby assigning different propagation weights to different paths of various heterogeneous graphs to better guide the propagation of the graph neural network. By distinguishing different source paths for control, the introduction of propagation noise can be avoided. At the same time, the propagation mechanism of the graph neural network is used to allow user nodes and resource nodes to "feel" the structural information of the graph and gradually acquire information about their neighborhood (or multi-hop neighborhood), thereby learning richer graph structure capabilities and improving their representation capabilities.

[0102] In one exemplary embodiment, step S122, based on the propagation weights and the initial embedding representations of each node, performs a multi-layer joint propagation operation of a graph neural network on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource. This includes: for any object among multiple users and multiple resources, determining the object embedding representation in each heterogeneous graph at each layer; fusing the object embedding representations in each heterogeneous graph to obtain the single-layer embedding representation of the object at each layer; and performing a weighted average of the single-layer embedding representations obtained at each layer to obtain the target embedding representation of the object.

[0103] Specifically, when only behavioral graphs are available, the embedding representation determined by each layer based on the behavioral graph is the single-layer embedding representation of that layer. However, when performing multi-layer joint propagation on multiple heterogeneous graphs based on graph neural networks, node embedding representations under three different heterogeneous graphs are determined in each layer. In this case, the embedding representations of the three heterogeneous graphs can be weighted and fused to obtain the single-layer embedding representation for each user and each resource. Finally, for each user, the target embedding representation for that user is obtained by weighted averaging of the single-layer embedding representations. Similarly, for each resource, the target embedding representation for that resource is obtained by weighted averaging of the single-layer embedding representations.

[0104] For any object among all users and resources, the joint propagation process of the graph neural network can be represented as:

[0105] (3)

[0106] in, The corresponding user and resource behavior edges in the behavior graph; Corresponding modal similarity edges in the modality diagram (as shown in the text-image distance calculation); Corresponding to knowledge edges in a knowledge graph built based on entity relationships; This represents the trainable or policy-defined path weight factor, which is the weight used to weight the three heterogeneous graphs.

[0107] In some embodiments, a Transformer can be used instead of formula (3) to fuse the embedding representations of the three heterogeneous graphs. In some embodiments, an edge weight training mechanism can be used to replace the static path factor. That is, using edge type and edge attributes as input features, training an edge weight predictor instead of a fixed edge weight predictor. .

[0108] The final node embedding aggregation can be represented as:

[0109] (4)

[0110] in, This represents the target embedding representation. This represents the embedding representation obtained from the behavior graph in each layer. This represents the embedding representation obtained from the modality diagram in each layer. This represents the embedding representation obtained from the knowledge graph in each layer, and the Fuse function is the fusion function.

[0111] The specific implementation of formula (4) includes two steps: First, in each layer, the embedding representations obtained from the three types of heterogeneous graphs are fused to obtain the single-layer embedding representation of each layer; Second, the single-layer embedding representations of each layer are weighted and averaged to obtain the target embedding representation.

[0112] For example, suppose we have the following interactive data:

[0113] user: ;

[0114] resource: ;

[0115] Interaction log: Clicked ; Clicked ;

[0116] With users For example, determine its target embedding representation. The process is as follows:

[0117] First layer: In the behavior graph, the user Embedded representation Based on resources The initial embedding representation is obtained by aggregation, that is, by performing initial embedding aggregation.

[0118] In modal diagrams, users Embedded representation Based on and It is obtained by aggregating the modal features of modally similar neighboring nodes. For example, and If the text features are similar, then... Text features embedded into the user The embedded representation, i.e., forming the propagation path: .

[0119] In a knowledge graph, users Embedded representation Based on and It is obtained by aggregating features from neighboring nodes that share common attribute information. For example, and If author A is shared, then there is a propagation path: .

[0120] Will , , The three results are weighted and summed using formula (3) to obtain the user's result. Single-layer embedding representation in the first layer: .

[0121] Second layer: In the behavior graph, the user Embedded representation Embedded interactive resources The interests of other users associated with it.

[0122] In modal diagrams, users Embedded representation Embedded and The modal characteristics of the neighboring nodes of modally similar neighboring nodes. For example, and If the image features are similar, then the embedding is propagated in the first layer. Based on the text features, embed again The image features that form the propagation path: .

[0123] In a knowledge graph, users Embedded representation Based on and It is obtained by aggregating the features of neighboring nodes that share common attribute information. For example, and Shared author A, and and Both belong to the "retro style" Also in retro style, there are .

[0124] Will , , The three results are weighted and summed using formula (3) to obtain the user's result. Single-layer embedding representation in the second layer: .

[0125] Finally, for users Single-layer embedding representation in the first layer Compared with single-layer embedding representation in the second layer Perform a weighted average to obtain the user Target embedding representation .

[0126] In some embodiments, meta-path sampling or path importance assessment strategies can also be used to prune and optimize the propagation path in order to reduce propagation costs.

[0127] In some embodiments, a graph neural network that natively supports multiple edge types can be used instead of a general graph neural network.

[0128] The multi-source path fusion propagation mechanism proposed in this embodiment can enhance semantic consistency propagation and avoid the bias caused by relying solely on structural edges by aggregating the propagation information of three types of heterogeneous graphs at the node level and learning an embedded representation containing collaborative semantics of behavior, resources, and knowledge.

[0129] In an exemplary embodiment, step S121 assigns propagation weights between two nodes with a connection relationship in each heterogeneous graph through a preset propagation control function, including: for modal graphs and knowledge graphs in each heterogeneous graph, assigning propagation weights between two nodes with a connection relationship through a preset propagation control function in a manner that the propagation weights are positively correlated with semantic relevance.

[0130] In behavioral graphs, whether a user has interacted with a resource is an objective result. However, in modal graphs and knowledge graphs, because these heterogeneous graphs are not generated by interactive behavior, they are prone to the propagation of "erroneous semantics," that is, irrelevant content is spread in, polluting the originally highly semantically consistent representation. Therefore, it is necessary to control its propagation weight, that is, to assign propagation weight to two nodes with a connection relationship in a way that the propagation weight is positively correlated with semantic relevance. That is, the higher the relevance, the higher the propagation weight, and the lower the relevance, the lower the propagation weight.

[0131] In this embodiment, for the modal graphs and knowledge graphs in each heterogeneous graph, propagation weights are assigned between two nodes that have a connection relationship, in accordance with the principle that propagation weights are positively correlated with semantic relevance. This makes the embedded representation of the obtained node closer to the "semantically consistent" neighboring node, while the influence of unrelated neighboring nodes is suppressed or filtered out. This enables the graph neural network to be guided to propagate along the semantically consistent path during propagation.

[0132] In one exemplary embodiment, the method further includes: determining a target embedding representation based on its modal graph and knowledge graph for a user among a plurality of users who has not interacted with any resource, or a resource among a plurality of resources that has not been interacted with by any user.

[0133] Specifically, for users without interactive data, weak connections can be established with semantic entities in the knowledge graph based on the interest tags selected during registration, explicit preferences, or the system's default initialization embedding. Simultaneously, connection edges can be established with similar resources in the modal graph through embeddings of popular resources, providing cold-start recommendations to users from a resource perspective. In this case, the propagation formula (3)... Without participating in the dissemination, the information will be disseminated through... and transfer.

[0134] For resources without interactive data, connections are established with other resources based on the modal features of the resource in the modal graph, and connections are established with other resources based on the attribute information of the resource in the knowledge graph. Similarly, in the propagation formula (3), Without participating in the dissemination, the information will be disseminated through... and transfer.

[0135] The method proposed in this embodiment can still utilize modal graphs and knowledge graphs for graph propagation and recall even in the case of no user or resource interaction data, i.e., cold start, thereby completing the recall capability and possessing a large degree of cold start adaptability.

[0136] It should be noted that the three types of heterogeneous graphs proposed in this application can each have their adjacency matrices and propagation outputs constructed and stored independently; the graph neural network propagation of each type of heterogeneous graph can be completed in parallel, and finally aggregated through the Fuse function; in engineering implementation, it can be deployed in the manner of "behavioral graph main path + pluggable supplementary graph module (modal graph / knowledge graph)". In other words, this application can adopt a modular structure, support decoupled deployment with existing graph retrieval systems, and has the advantages of easy integration and deployment.

[0137] In one exemplary embodiment, in order to facilitate understanding of the embodiments of this disclosure by those skilled in the art, the following will be described in conjunction with the accompanying drawings. Figure 7 To further explain this method, such as Figure 7 The diagram shown illustrates a resource retrieval method according to another embodiment. This embodiment includes the following steps:

[0138] (1) Construct multiple heterogeneous graphs based on the historical interaction datasets between multiple users and multiple resources.

[0139] Multiple heterogeneous graphs include a behavioral graph representing user-resource interaction records, a modal graph representing the relationships of various resources under different modal features, and a knowledge graph representing the relationships of various resources under preset attribute information.

[0140] (2) By using a preset propagation control function, propagation weights are assigned between two nodes that are connected in each heterogeneous graph; wherein, in the propagation control function, the gating network corresponding to each heterogeneous graph is different.

[0141] Specifically, for modal graphs and knowledge graphs in each heterogeneous graph, propagation weights are assigned between two nodes with a connection relationship through a preset propagation control function, in accordance with the positive correlation between propagation weights and semantic relevance.

[0142] (3) For any object among multiple users and multiple resources, in each layer, determine the object embedding representation of the object in each heterogeneous graph.

[0143] (4) The object embedding representations in each heterogeneous graph are fused to obtain the single-layer embedding representation of the object in each layer.

[0144] (5) Perform a weighted average of the single-layer embedding representations obtained in each layer to obtain the target embedding representation of the object.

[0145] Specifically, for users who have not interacted with any resources among multiple users, or resources that have not been interacted with by any user among multiple resources, the target embedding representation is determined based on their modal graph and knowledge graph.

[0146] (6) For each resource, calculate the recall score for the resource based on the target embedding representation of the resource and the target embedding representation of each user;

[0147] (7) Based on the recall score of each resource, determine the recall resources from the plurality of resources.

[0148] This method integrates user behavior graphs, resource modality graphs, and knowledge graphs into a unified heterogeneous graph structure, breaking the limitation of traditional recommendation systems that rely solely on user-resource behavior to construct the graph. By introducing a path-weighted propagation mechanism, the system can dynamically adjust structural paths from different sources (such as behavior paths, modality paths, and knowledge paths), guiding information to propagate preferentially along semantically consistent directions, effectively alleviating the problems of structural fragmentation and semantic disconnect. Simultaneously, this method proposes a cold-start "graph supplementation" mechanism, allowing users without behavior or new resource nodes to participate in propagation through modal features and knowledge entities, improving the system's recall and coverage capabilities in cold-start and long-tail scenarios. Finally, through a multi-source path fusion strategy, the propagated information from different types of graphs is uniformly aggregated, achieving richer and semantically consistent representation learning for nodes, significantly improving recommendation accuracy and system stability. Furthermore, this solution possesses good modularity and scalability, and can be flexibly integrated into existing graph recall systems.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0150] It is understood that the same / similar parts between the various embodiments of the methods described above in this specification can be referred to each other. Each embodiment focuses on the differences from other embodiments, and relevant parts can be referred to the description of other method embodiments.

[0151] Based on the same inventive concept, this disclosure also provides a resource recall apparatus for implementing the resource recall method described above.

[0152] Figure 8 This is a structural block diagram illustrating a resource retrieval device according to an exemplary embodiment. (Refer to...) Figure 8 The device includes:

[0153] Heterogeneous graph construction unit 810 is configured to construct multiple heterogeneous graphs based on a historical interaction dataset between multiple users and multiple resources. The multiple heterogeneous graphs include a behavioral graph representing the interaction records between users and resources, a modal graph representing the relationship between various resources under different modal features, and a knowledge graph representing the relationship between various resources under preset attribute information.

[0154] The information propagation unit 820 is configured to perform the determination of the target embedding representation of each user and the target embedding representation of each resource based on the connection relationships between nodes in each heterogeneous graph and the initial embedding representation of each node.

[0155] The score calculation unit 830 is configured to perform a recall score for each resource based on the target embedding representation of the resource and the target embedding representation of each user.

[0156] Resource filtering unit 840 is configured to perform a recall score based on each resource to identify recall resources from a pool of resources.

[0157] In an exemplary embodiment, each modal feature corresponds to a modal graph; the heterogeneous graph construction unit 810 is further configured to perform the following: for each modal feature, obtain the similarity between every two resources among multiple resources under the modal feature; and establish connection edges between the nodes corresponding to two resources with similarity higher than a threshold, using each resource as a node, to obtain the modal graph under the modal feature.

[0158] In an exemplary embodiment, there are multiple preset attribute information; the heterogeneous graph construction unit 810 is further configured to perform the determination of the association between each resource and each attribute information, as well as the association between each attribute information; using each resource and each attribute information as nodes, establish connection edges between related resources and attribute information, and establish connection edges between related attribute information, to obtain a knowledge graph.

[0159] In an exemplary embodiment, the information propagation unit 820 is further configured to execute a preset propagation control function to assign propagation weights between two nodes that are connected in each heterogeneous graph; wherein, in the propagation control function, the gating networks corresponding to each heterogeneous graph are different; based on the propagation weights and the initial embedding representations of each node, a multi-layer joint propagation operation of a graph neural network is performed on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource.

[0160] In an exemplary embodiment, the information propagation unit 820 is further configured to perform the following actions for any one of multiple users and multiple resources: in each layer, determine the object embedding representation of the object in each heterogeneous graph; fuse the object embedding representations in each heterogeneous graph to obtain a single-layer embedding representation of the object in each layer; and perform a weighted average of the single-layer embedding representations obtained in each layer to obtain the target embedding representation of the object.

[0161] In an exemplary embodiment, the information propagation unit 820 is further configured to perform propagation control functions on modal graphs and knowledge graphs in each heterogeneous graph, assigning propagation weights to two nodes with a connection relationship in a manner in which propagation weights are positively correlated with semantic relevance.

[0162] In one exemplary embodiment, the information dissemination unit 820 is further configured to determine a target embedding representation based on its modal graph and knowledge graph for a user among a plurality of users who has not interacted with any resource, or a resource among a plurality of resources that has not been interacted with by any user.

[0163] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0164] Figure 9 This is a block diagram illustrating an electronic device 900 for implementing a resource retrieval method according to an exemplary embodiment. For example, the electronic device 900 can be a server. (Refer to...) Figure 9 The electronic device 900 includes a processing component 920, which further includes one or more processors, and memory resources represented by memory 922 for storing instructions, such as application programs, that can be executed by the processing component 920. The application programs stored in memory 922 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 920 is configured to execute instructions to perform the methods described above.

[0165] Electronic device 900 may further include: a power supply component 924 configured to perform power management of electronic device 900, a wired or wireless network interface 926 configured to connect electronic device 900 to a network, and an input / output (I / O) interface 928. Electronic device 900 can operate on an operating system stored in memory 922, such as Windows 99erver, Mac O9 X, Unix, Linux, FreeB9D, or similar.

[0166] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 922 including instructions, which can be executed by a processor of an electronic device 900 to perform the above-described method. The storage medium may be a computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0167] In an exemplary embodiment, a computer program product is also provided, the computer program product including instructions that can be executed by a processor of an electronic device 900 to perform the above method.

[0168] It should be noted that the above-mentioned apparatus, server, computer-readable storage medium, computer program product, etc., may also include other implementation methods according to the description of the method embodiments. For specific implementation methods, please refer to the description of the relevant method embodiments, which will not be elaborated here.

[0169] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0170] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A resource recall method, characterized in that, include: Based on the historical interaction dataset between multiple users and multiple resources, multiple heterogeneous graphs are constructed. The multiple heterogeneous graphs include a behavioral graph representing the interaction records between users and resources, a modal graph representing the relationship between each resource under different modal features, and a knowledge graph representing the relationship between each resource under preset attribute information. By using a preset propagation control function, propagation weights are assigned between two nodes that are connected in each heterogeneous graph; wherein, the propagation weights in the modal graph and the knowledge graph are determined in a way that the propagation weights are positively correlated with semantic relevance; Based on the propagation weights and the initial embedding representations of each node, a multi-layer joint propagation operation of a graph neural network is performed on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource. Specifically, this includes: for any object among the multiple users and multiple resources, in each layer, determining the object embedding representation of the object in each heterogeneous graph; fusing the object embedding representations in each heterogeneous graph to obtain the single-layer embedding representation of the object in each layer; and performing a weighted average of the single-layer embedding representations obtained in each layer to obtain the target embedding representation of the object. During the multi-layer joint propagation process, each heterogeneous graph in each layer will share the embedding representation of each object. The embedding representation of each object in the current layer is calculated through the connection relationships between nodes in each heterogeneous graph, the path weight factors corresponding to each heterogeneous graph, and the embedding representation of the object in the previous layer. For each resource, a recall score for that resource is calculated based on the target embedding representation of the resource and the target embedding representation of each user. Based on the recall score of each resource, the recalled resources are determined from the plurality of resources.

2. The method according to claim 1, characterized in that, Each modal feature corresponds to a modality graph; the process of constructing the modality graph based on the historical interaction dataset includes: For each modal feature, the similarity between every two resources among the multiple resources under that modal feature is obtained; Using each resource as a node, establish connection edges between nodes corresponding to two resources with similarity higher than a threshold to obtain the modality graph under this modality feature.

3. The method according to claim 1, characterized in that, There are multiple preset attribute information; the process of constructing the knowledge graph based on the historical interaction dataset includes: Determine the relationship between each resource and each attribute information, as well as the relationship between each attribute information; Using each resource and each attribute information as nodes, connection edges are established between related resources and attribute information, and between related attribute information, to obtain the knowledge graph.

4. The method according to claim 1, characterized in that, In the propagation control function, the gating networks corresponding to each heterogeneous graph are different.

5. The method according to claim 1, characterized in that, The method further includes: For users who have not interacted with any resources among the multiple users, or resources that have not been interacted with by any user among the multiple resources, the target embedding representation is determined based on their modal graph and knowledge graph.

6. A resource recall device, characterized in that, include: The heterogeneous graph building unit is configured to construct multiple heterogeneous graphs based on a historical interaction dataset between multiple users and multiple resources. The multiple heterogeneous graphs include a behavioral graph representing the interaction records between users and resources, a modal graph representing the relationship between various resources under different modal features, and a knowledge graph representing the relationship between various resources under preset attribute information. The information propagation unit is configured to execute a preset propagation control function to assign propagation weights between two nodes that are connected in each heterogeneous graph; based on the propagation weights and the initial embedding representations of each node, a multi-layer joint propagation operation of a graph neural network is performed on each heterogeneous graph to obtain the target embedding representation of each user and the target embedding representation of each resource; wherein, the propagation weights in the modal graph and the knowledge graph are determined in a way that the propagation weights are positively correlated with semantic relevance; The information propagation unit is further configured to perform the following steps for any one of the multiple users and multiple resources: in each layer, determine the object embedding representation of the object in each heterogeneous graph; fuse the object embedding representations in each heterogeneous graph to obtain the single-layer embedding representation of the object in each layer; and perform a weighted average of the single-layer embedding representations obtained in each layer to obtain the target embedding representation of the object. During multi-layer joint propagation, each heterogeneous graph in each layer will share the embedding representation of each object. The embedding representation of each object in the current layer is calculated using the connection relationships between nodes in each heterogeneous graph, the path weight factors corresponding to each heterogeneous graph, and the embedding representation of the object in the previous layer. The scoring unit is configured to perform a recall score for each resource based on the target embedding representation of the resource and the target embedding representation of each user. The resource filtering unit is configured to perform a recall score based on each resource to determine the recall resources from the plurality of resources.

7. The apparatus according to claim 6, characterized in that, Each modal feature corresponds to a modal graph; the heterogeneous graph construction unit is also configured to perform the following: for each modal feature, obtain the similarity between every two resources among the multiple resources under the modal feature; and establish connection edges between the nodes corresponding to two resources with similarity higher than a threshold, using each resource as a node, to obtain the modal graph under the modal feature.

8. A server, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the resource recall method as described in any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the server, the server is enabled to perform the resource retrieval method as described in any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the resource recall method as described in any one of claims 1 to 5.

Images