Resource recall method and device, server and storage medium

Abstract

The invention relates to a resource recall method and device, a server and a storage medium, and the method comprises the steps: constructing a plurality of heterogeneous graphs based on a historical interaction data set between a plurality of users and a plurality of resources; the plurality of heterogeneous graphs comprise a behavior graph for representing interaction records of users and resources, a modal graph for representing the relationship of each resource under different modal characteristics, and a knowledge graph for representing the relationship of each resource under preset attribute information; determining a target embedding representation of each user and a target embedding representation of each resource based on a connection relationship between nodes in each heterogeneous graph and the initial embedding representation of each node; for each resource, calculating a recall score for the resource based on the target embedding representation of the resource and the target embedding representation of each user; and based on the recall score of each resource, determining a recalled resource from the plurality of resources. According to the method, the high-quality recall capability can be still kept in a data sparse and cold start environment.

Description

Technical Field

[0001] The present disclosure relates to the field of recommendation technology, and in particular to a resource recall method, device, server, storage medium, and program product. Background Art

[0002] In the recommendation field, there are two main methods for resource recall in recommendation systems. One is to apply graph neural networks for recall modeling, specifically by constructing a user-item interaction graph (User-Item Graph) and using simplified graph convolution operations to learn node representations, thereby achieving collaborative filtering recommendations between users and resources. The other is multimodal recommendation, which initializes each resource node as a representation that integrates multimodal features such as images, text, and labels, and then performs graph neural network propagation operations based on the user-resource interaction graph. Modal information does not participate in the edge construction or path regulation of the graph structure itself.

[0003] However, both of these methods rely solely on the interaction between users and resources during the dissemination process. 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 a broken graph structure, making it impossible to establish a dissemination path, and affecting recall capabilities. Summary of the Invention

[0004] The present disclosure provides a resource recall method, device, server, storage medium, and program product to at least address the problem in related technologies of poor recall capabilities for new users with no behavioral records or in situations where behavioral data is sparse. The technical solutions of the present disclosure are as follows:

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

[0006] Based on historical interaction datasets between multiple users and multiple resources, multiple heterogeneous graphs are constructed; the multiple heterogeneous graphs include a behavior graph representing the interaction records between users and resources, a modal graph representing the relationship between each resource under different modal characteristics, and a knowledge graph representing the relationship between each resource under preset attribute information;

[0007] Based on the connection relationship between the 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, calculate a recall score for the resource based on the target embedding representation of the resource and the target embedding representation of each user;

[0009] A recalled resource is determined from the plurality of resources based on the recall score of each resource.

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

[0011] For each modal feature, obtaining the similarity between every two resources in the plurality of resources under the modal feature;

[0012] Taking each resource as a node, a connecting edge is established between the nodes corresponding to two resources with a similarity higher than a threshold, and a modal graph under the modal feature is obtained.

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

[0014] Determine the association between each resource and each attribute information, and the association between each attribute information;

[0015] With each resource and each attribute information as a node, connecting edges are established between related resources and attribute information, and connecting edges are established between related attribute information to obtain the knowledge graph.

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

[0017] Using a preset propagation control function, a propagation weight is assigned between two nodes in a connection relationship in each heterogeneous graph; wherein, in the propagation control function, each heterogeneous graph corresponds to a different gating network;

[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 a target embedding representation of each user and a target embedding representation of each resource.

[0019] In an exemplary embodiment, 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 representation of each node to obtain a target embedding representation of each user and a target embedding representation of each resource includes:

[0020] For any object among the multiple users and the multiple resources, determining, in each layer, an object embedding representation of the object in each heterogeneous graph;

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

[0022] The single-layer embedding representation obtained at each layer is weighted averaged to obtain the target embedding representation of the object.

[0023] In an exemplary embodiment, assigning a propagation weight between two nodes having a connection relationship in each heterogeneous graph by using a preset propagation control function includes:

[0024] For the modal graphs and knowledge graphs in each heterogeneous graph, a preset propagation control function is used to assign propagation weights between two nodes with a connection relationship in a way that the propagation weights are positively correlated with the semantic relevance.

[0025] In an exemplary embodiment, the method further comprises:

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

[0027] According to a second aspect of an embodiment of the present disclosure, there is provided a resource recall device, comprising:

[0028] A heterogeneous graph construction unit is configured to construct a plurality of heterogeneous graphs based on a historical interaction dataset between a plurality of users and a plurality of resources; the plurality of heterogeneous graphs include a behavior graph representing interaction records between users and resources, a modal graph representing relationships between resources under different modal features, and a knowledge graph representing relationships between resources under preset attribute information;

[0029] An information propagation unit is configured to determine a target embedding representation of each user and a target embedding representation of each resource based on the connection relationship between nodes in each heterogeneous graph and the initial embedding representation of each node;

[0030] a score calculation unit configured to calculate, for each resource, a recall score for the resource based on a target embedding representation of the resource and a target embedding representation of each user;

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

[0032] In an exemplary embodiment, each modal feature corresponds to a modal graph; the heterogeneous graph construction unit is further configured to execute, for each modal feature, obtaining the similarity between each two resources among the multiple resources under the modal feature; taking each resource as a node, establishing a connecting edge between the nodes corresponding to two resources whose similarity is higher than a threshold, and obtaining 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 determine 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 a node, establishing connecting edges between associated resources and attribute information, and establishing connecting edges between associated 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 a propagation weight between two nodes with a connection relationship in each heterogeneous graph; wherein, in the propagation control function, the gating network corresponding to each heterogeneous graph is different; based on the propagation weight and the initial embedding representation of each node, a multi-layer joint propagation operation of the 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 propagation unit is further configured to perform, for any one of the multiple users and the multiple resources, determining the object embedding representation of the object in each heterogeneous graph in each layer; fusing the object embedding representations in each heterogeneous graph to obtain a single-layer embedding representation of the object in each layer; and performing weighted averaging on 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 execute, for the modal graph and knowledge graph in each heterogeneous graph, a preset propagation control function, and assign propagation weights between two nodes with a connection relationship in a manner such that the propagation weights are positively correlated with the semantic relevance.

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

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

[0039] processor;

[0040] a memory for storing instructions executable by the processor;

[0041] The processor is configured to execute the instructions to implement any of the above methods.

[0042] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute any of the methods described above.

[0043] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, wherein the computer program product includes instructions, and when the instructions are executed by a processor of a server, the server is enabled to execute any of the methods described above.

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

[0045] Based on the historical interaction dataset between multiple users and multiple resources, the present invention constructs multiple heterogeneous graphs such as a behavior 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. Based on each heterogeneous graph, the target embedding representation of each user and the target embedding representation of each resource are jointly determined. Further, based on the target embedding representation of the resource and the target embedding representation of each user, the recall score of the resource is calculated, thereby determining the recalled resource from multiple resources. This multi-source information joint recall method that combines the behavior graph, modal graph and knowledge graph structure no longer relies on "a certain" single relationship to connect users and resources, but uses three types of edges: user behavior, modal similarity and knowledge entity co-occurrence to construct a heterogeneous graph. It can achieve richer and more semantically consistent representation learning for nodes, so that high-quality recall capabilities can be maintained in data sparse and cold start environments, significantly improving recommendation accuracy and system stability, and solving the problems of behavior sparsity, resource cold start and semantic disconnection in complex recommendation scenarios.

[0046] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0048] Figure 1 The figure is a flowchart of a resource recall method according to an exemplary embodiment.

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

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

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

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

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

[0054] Figure 7 The figure is a flowchart of a resource recall method according to another exemplary embodiment.

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

[0056] Figure 9 The figure is a block diagram of a server according to an exemplary embodiment. DETAILED DESCRIPTION

[0057] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0058] It should be noted that the implementations described in the following exemplary embodiments do not represent all implementations consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of this disclosure, as detailed in the appended claims. It should also be noted that all 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 and analysis, etc.) referred to in this disclosure are authorized by the user or have been fully authorized by all parties.

[0059] Understandably, recommendation systems in related technologies often rely on user-resource interactions to construct graph structures and implement node embedding generation through graph neural networks. However, this type of approach faces three difficulties in actual platforms: (1) Behavior sparsity: a large amount of long-tail content lacks effective interaction with new users, resulting in a broken graph structure and an inability to establish a propagation path; (2) Modal heterogeneity: content often has multimodal attributes such as images, text, and audio, and a single structure cannot carry all the information; (3) Semantic disconnection: a large amount of semantically related content exists but has no direct structural connection, such as "same type of work" or "same author", which is difficult to model in a 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 (behavior graph + modal graph + knowledge graph) and perform joint propagation control during the graph neural network propagation process to enhance 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. Instead, it uses three types of edges: user behavior, modal similarity, and knowledge entity co-occurrence to construct a supergraph. Through a designed propagation mechanism, it controls how these three types of edges participate in the graph neural network, thereby achieving structural enhancement and semantic completion.

[0061] In an exemplary embodiment, Figure 1 As shown, a resource recall method is provided. This embodiment uses the method applied to a server as an example. The server can be implemented as an independent 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 data sets between multiple users and multiple resources; the multiple heterogeneous graphs include a behavior 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.

[0063] The resource represents a content entity that can be recommended to the user. For example, the resource may be a video.

[0064] The behavior graph is formed by establishing interaction edges between users and resources based on user clicks, likes, and browsing behaviors. The behavior graph has two types of nodes: user nodes and resource nodes. It is established between the user node and the resource node, and only when the user and resources It is established when there is interaction (such as click, purchase). Assume the following interaction data: User: ,resource: ;Interaction record: Clicked ; Clicked , then the constructed behavior diagram is as follows Figure 2 As shown, in the user and resources There is a connection edge between the user and resources There are connecting edges between them.

[0065] Among them, the modal graph is a modal similarity connection established between resources based on the modal features of resources (such as images, text, etc.). The nodes in the modal graph are resource nodes. A modal graph is formed under each modality. For example, a modal graph can be generated under text features, and a modal graph can also be generated under image features. Figure 3 As shown, Figure 3 Figures (a) and (b) in the figure represent the modal graphs under text features and image features respectively. Figure 3 Figure (a) shows the resources and resources ,resource ,resource The text features of the resources are similar. and resources ,resource The text features are similar; (b) The graph represents the resource and resources ,resource The image features are similar, resources and resources The image features are similar.

[0066] The knowledge graph represents a semantic network built based on the attribute information of resources (such as tags, authors, categories, styles, etc.). The nodes in the knowledge graph include resource nodes and entity nodes corresponding to each attribute information. Resources are indirectly connected by 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 They can be indirectly connected through the shared label Y, and there is an entity connection between category Y and author B.

[0067] It can be understood that traditional recall methods only use user-resource interaction edges to construct graph structures and cannot introduce modal information such as images and text, resulting in poor recall performance on cold-start resources or users with no interactive behavior. Therefore, when constructing heterogeneous graphs in this step, in addition to constructing behavioral graphs, modal graphs and knowledge graphs are also introduced to jointly represent the relationship between users and resources, breaking the limitation of traditional recommendation systems that only rely on user-resource behavior to construct graph structures.

[0068] In step S120 , based on the connection relationship between the 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] Among them, user embeddings and item embeddings are numerical representations that map users and resources into low-dimensional vector spaces. These embeddings can capture the latent semantics of user preferences and resource characteristics.

[0070] In the specific implementation, the target embedding representation (or final embedding representation) of users and resources is determined through the information propagation mechanism of the graph neural network. Specifically, the information propagation operation of the graph neural network mainly includes four steps: initializing node embedding, neighbor aggregation, multi-layer propagation, and generating the final embedding. Among them, initializing node embedding is to initialize the initial embedding of each node in each heterogeneous graph. For example, the initial embedding representation of a user can be the user's interests, age, occupation, etc., and the embedding representation of a resource can be the resource's characteristics (such as content, description, etc.). Neighbor aggregation represents that in each layer of the graph neural network, the node updates its own representation by aggregating information from neighboring nodes. The embedding representation of users and resources in each layer of multi-layer propagation is obtained by combining the embedding representation obtained in the previous layer. Finally, the embedding representations of users and resources in all layers are weighted 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 to the liked resources, and resources with similar attribute information to the liked resources. Even if some users and resources have no direct interaction, indirect associations can be established through the graph structure, alleviating the sparsity problem of interaction data. At the same time, dynamic interest diffusion can be achieved, allowing user interest to spread from directly interacted 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] Among them, the recall score can be used to measure the matching degree between users and resources, thereby screening out recall resources with higher matching degree from massive candidate resources.

[0074] For example, for each resource, we can calculate the inner product of the target embedding representation of each user and the target embedding representation of the resource to obtain multiple scores for the resource. The average of the scores is calculated to obtain the recall score of the resource. This process can be expressed as:

[0075] (1)

[0076] in, Represents a user Resources The recall score of Represents a user The target embedding representation is Represents a resource The target embedding representation of

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

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

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

[0080] In a specific implementation, a higher recall score indicates a greater probability that the user is interested in the resource, and a higher recommendation degree. Therefore, the top N resources with higher recall scores can be determined from multiple resources in descending order of recall scores as the recalled resources.

[0081] In the above-mentioned resource recall method, based on a dataset of historical interactions between multiple users and multiple resources, multiple heterogeneous graphs are constructed, including a behavior graph representing the interaction records between users and resources, a modal graph representing the relationships between resources under different modal characteristics, and a knowledge graph representing the relationships between resources under preset attribute information. Based on these heterogeneous graphs, a target embedding representation for each user and a target embedding representation for each resource are jointly determined. Furthermore, 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, thereby determining the recalled resource from the multiple resources. This multi-source information joint recall method, which combines behavior graphs, modal graphs, and knowledge graph structures, no longer relies on a single relationship to connect users and resources. Instead, it constructs a heterogeneous graph using three types of edges: user behavior, modal similarity, and knowledge entity co-occurrence. This allows for richer and more semantically consistent representation learning of nodes, maintaining high-quality recall capabilities even in data sparseness and cold start environments, significantly improving recommendation accuracy and system stability, and addressing the issues of behavior sparsity, resource cold start, and semantic disconnection in complex recommendation scenarios.

[0082] It can be understood that a modal graph connects different resources through modalities, and the connection relationships between resources under different modal features are different. Therefore, it is necessary to construct a corresponding modal graph for each modal feature to represent the connection relationships between resources under different modal features. In an exemplary embodiment, in step S110, based on the historical interaction dataset, the process of constructing a modal graph includes: for each modal feature, obtaining the similarity between each two resources in a plurality of resources under the modal feature; using each resource as a node, establishing a connecting edge between the nodes corresponding to two resources with a similarity greater than a threshold, and obtaining a modal graph under the modal feature.

[0083] Specifically, multiple modal features, such as text features and image features, are identified for constructing a modal graph. For each modal feature, the similarity between each pair of resources under that modal feature is calculated. Connecting edges are established between the nodes corresponding to two resources with similarities above a threshold, thereby forming a modal graph under that modal feature.

[0084] For example, assuming the resource and resources ,resource ,resource If the similarity of the text features of and resources ,resource ,resource Establish connection edges between them, resources and resources ,resource If the similarity of the text features of and resources ,resource A connection is also established between them, and the resulting modal diagram is as follows Figure 3 As shown in Figure (a) in . For example, assuming that the resource and resources ,resource If the similarity of the image features is higher than the threshold, the and resources ,resource Establish connection edges between them; resources and resources If the similarity of the image features is higher than the threshold, the and resources Create connecting edges between them.

[0085] In this embodiment, through the modal graphs of each resource under multiple modal features, the association relationship between resources is established based on the multimodal features of the resources, so that semantically consistent path selection can be guided based on the modal features in the future without relying solely on the behavioral interaction between users and resources. This can improve the representation capabilities of users and resources, thereby achieving high-quality resource recall.

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

[0087] Specifically, when constructing a knowledge graph, each resource and each piece of attribute information is used as a node. After determining the association between each resource and each piece of attribute information, a connecting edge is established between each resource and each piece of attribute information. If there is also a correlation between the individual pieces of attribute information, a connecting edge is established between the related attribute information.

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

[0089] In some embodiments, given the varying strengths of the correlations between different attribute information, the connections between attribute information can be either weak or strong. For example, if author A's works primarily feature style B, with some occasionally featuring style C, then author A and style B are strongly connected, while author A and style C are weakly connected.

[0090] In this embodiment, by constructing a knowledge graph based on the attribute information of each resource, semantic association modeling is achieved: for example, connecting edges between semantically similar resources such as "same author" and "same style." By modeling the semantic structural relationships between complex resources, this method can improve the richness and generalization of resource embedding representations, thereby enhancing resource recall quality.

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

[0092] Step S121, assigning a propagation weight between two nodes having a connection relationship in each heterogeneous graph through a preset propagation control function;

[0093] In step S122, based on the propagation weights and the initial embedding representations of each node, a multi-layer joint propagation operation of the 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.

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

[0095] (2)

[0096] in, and is the embedding representation of two nodes, Path types for different heterogeneous graphs In the gating mechanism, the core task of the MLP is to map input features into gating weights. It can be understood that the propagation weight (or attention) of the edge connecting two nodes determines the importance distribution of neighboring nodes during information propagation. In this embodiment, different heterogeneous graphs correspond to different gating networks, and different propagation weights are assigned to the edges connecting the heterogeneous graphs. This controls the propagation weights of different heterogeneous graphs and ensures the accuracy and effectiveness of the propagation process.

[0097] In some embodiments, a Graph Attention Network (GAT) can also be used to calculate attention scores for the three types of edges in the three types of heterogeneous graphs, using these scores as propagation weights to dynamically adjust the propagation paths. It should be noted that the GAT learns attention for each edge independently, without distinguishing between edge types.

[0098] Specifically, the propagation operation of a graph neural network involves transferring and aggregating information based on the interactions 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 to leverage the graph structure to capture collaborative filtering signals (e.g., "similar users like similar resources"). The propagation layer of a graph neural network typically consists of 1-3 layers, and propagation within each layer includes three key operations: message passing, aggregation, and updating. Message passing involves 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 its edges (if any). Aggregation involves each node aggregating messages from its neighbors into a comprehensive set of neighborhood information. Updates involve combining the node's own features from the previous round with the aggregated neighborhood information to generate new features for the node.

[0099] More specifically, in the first layer, nodes only aggregate information of direct neighbor nodes; in the second layer, nodes aggregate information of neighbor nodes of neighbor nodes; and so on. User nodes and resource nodes can capture the structure and characteristics of neighborhood nodes and realize hierarchical feature learning.

[0100] For example, assume that in the first layer, node A aggregates the information of node B and node C, and node B aggregates the information of node D. Then, in the second layer, node A indirectly aggregates the information of node D.

[0101] In this embodiment, a propagation control function is designed, and different gating networks are designed for different types of heterogeneous graphs, so as to assign different propagation weights to different paths of various heterogeneous graphs, so as 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 obtain information about their neighborhood (or multi-hop neighborhood), so as to learn richer graph structure capabilities and improve their representation capabilities.

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

[0103] Specifically, when only a behavior graph is available, the embedding representation determined at each layer based on the behavior graph is the single-layer embedding representation of that layer. However, when performing multi-layer joint propagation on multiple heterogeneous graphs based on a graph neural network, node embedding representations for the three heterogeneous graphs are determined at each layer. In this case, the embedding representations of the three heterogeneous graphs can be weighted and fused to obtain a single-layer embedding representation for each user and a single-layer embedding representation for each resource. Finally, for each user, the target embedding representation for that user is obtained by taking a weighted average of the user's single-layer embedding representations. Similarly, for each resource, the target embedding representation for that resource is obtained by taking a weighted average of the resource's single-layer embedding representations.

[0104] For each user and any object in each resource, the joint propagation process of the graph neural network can be expressed as:

[0105] (3)

[0106] in, The behavior edges between users and resources in the corresponding behavior graph; Corresponding to the modal similarity edge in the modal graph (as calculated by graph and text distance); Corresponding to the knowledge edge in the knowledge graph built based on entity relationships; It represents the path weight factor that can be trained or set by the strategy, that is, the weight for weighting the three heterogeneous graphs.

[0107] In some embodiments, the Transformer can be used instead of formula (3) to fuse the embedding representations of the three heterogeneous graphs. In some embodiments, the static path factor can be replaced by the edge weight training mechanism. , that is, taking edge type and edge attribute as input features, training edge weight predictor instead of fixed .

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

[0109] (4)

[0110] in, represents the target embedding representation, represents the embedded representation based on the behavior graph in each layer, represents the embedding representation based on the modal graph in each layer, It represents the embedded representation based on the knowledge graph in each layer, and the Fuse function is the fusion function.

[0111] The specific implementation of formula (4) includes two steps: the first step is to fuse the embedding representations obtained from the three types of heterogeneous graphs in each layer to obtain the single-layer embedding representation of each layer; the second step is to perform weighted averaging on the single-layer embedding representation of each layer to obtain the target embedding representation.

[0112] For example, suppose there is the following interaction data:

[0113] user: ;

[0114] resource: ;

[0115] Interaction log: Clicked ; Clicked ;

[0116] User As an example, determine its target embedding representation The process is as follows:

[0117] First layer: In the behavior diagram, users Embedded representation of Resource-based The initial embedding representation of is obtained by aggregation, that is, initial embedding aggregation is performed.

[0118] In the modal diagram, the user Embedded representation of Based on The modal features of neighboring nodes with similar modalities are aggregated. For example, and If the text features of The text features are embedded into the user The embedded representation of , that is, forming a propagation path: .

[0119] In the knowledge graph, users Embedded representation of Based on The features of neighboring nodes with common attribute information are aggregated. For example, and Sharing author A, there are propagation paths: .

[0120] Will 、 、 The three results are weighted and summed using formula (3) to obtain user A single layer embedding representation at the first level: .

[0121] Second layer: In the behavior diagram, users Embedded representation of Embed interactive resources The interests of other users you are connected to.

[0122] In the modal diagram, the user Embedded representation of Embedded with The modal features of the neighboring nodes of the neighboring nodes with similar modalities. For example, and The image features are similar, so the embedding is propagated in the first layer Based on the text features, embed The image features form the propagation path: .

[0123] In the knowledge graph, users Embedded representation of Based on The features of neighbor nodes with common attribute information are aggregated. For example, and Shares author A, and and Both belong to the "retro style". It is also a retro style. .

[0124] Will 、 、 The three results are weighted and summed using formula (3) to obtain user The single-layer embedding representation at the second layer is: .

[0125] Finally, for users A single layer embedding representation at the first layer Compared with the single layer embedding representation in the second layer Perform weighted average to get user The target embedding representation .

[0126] In some embodiments, meta-path sampling or path importance evaluation strategies may also be used to perform pruning optimization of the propagation path to reduce the propagation cost.

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

[0128] The multi-source path fusion propagation mechanism proposed in this embodiment aggregates the propagation information of three types of heterogeneous graphs at the node level and learns an embedded representation that contains the semantics of behavior-resource-knowledge collaboration. This can enhance semantic consistency propagation and avoid the deviation caused by relying solely on structural connections.

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

[0130] In behavioral graphs, whether or not users interact with resources is an objective result. However, in modal graphs and knowledge graphs, since these heterogeneous graphs are not generated by interactive behaviors, they are prone to the propagation of "false semantics," where irrelevant content is introduced, contaminating the originally highly semantically consistent representation. Therefore, it is necessary to control the propagation weight. Specifically, a propagation weight is assigned to the connection between two nodes in a manner that is positively correlated with semantic relevance. 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, a propagation weight is assigned between two nodes with a connection relationship in a manner that the propagation weight is positively correlated with the semantic relevance, so that the resulting embedded representation of the node is closer to the "semantically consistent" neighboring node, and the influence of irrelevant neighboring nodes is suppressed or filtered out, so that when propagating, the graph neural network is guided to propagate preferentially along the semantically consistent path.

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

[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 embedding initialized by the system default. At the same time, connections can be established with the embeddings of similar resources to popular resources in the modal graph to provide users with cold start recommendations from a resource perspective. In this case, in the propagation formula (3), Do not participate in the dissemination, the information will pass 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), Do not participate in the dissemination, the information will pass and transfer.

[0135] The method proposed in this embodiment can use modal graphs and knowledge graphs for graph propagation and recall even when there is no interactive data between users or resources, that is, in the case of cold start, thereby completing the recall capability and having a high degree of cold start adaptability.

[0136] It should be noted that the three types of heterogeneous graphs proposed in this application can each be independently constructed and their adjacency matrices and propagation outputs stored. Graph neural network propagation for each type of heterogeneous graph can be completed in parallel, ultimately aggregated through a fuse function. In engineering implementation, this can be deployed using a "behavioral graph main path + pluggable graph supplementation modules (modal graph / knowledge graph)." This means that this application can adopt a modular structure, supporting decoupled deployment from existing graph recall systems and facilitating easy integration and deployment.

[0137] In an exemplary embodiment, in order to facilitate those skilled in the art to understand the embodiment of the present disclosure, the following will be combined with the attached Figure 7 To further illustrate this method, Figure 7 FIG. 1 is a flow chart of a resource recall method provided in another embodiment. In this embodiment, the method includes the following steps:

[0138] (1) Based on the historical interaction datasets between multiple users and multiple resources, multiple heterogeneous graphs are constructed.

[0139] The multiple heterogeneous graphs include a behavior 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.

[0140] (2) Through the preset propagation control function, a propagation weight is assigned between two nodes with a connection relationship in each heterogeneous graph; wherein, in the propagation control function, the gating network corresponding to each heterogeneous graph is different.

[0141] Among them, for the modal graph and knowledge graph in each heterogeneous graph, a preset propagation control function is used to assign propagation weights between two nodes with a connection relationship in a way that the propagation weights are positively correlated with the semantic relevance.

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

[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 weighted averaging on the single-layer embedding representations obtained at each layer to obtain the target embedding representation of the object.

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

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

[0147] (7) Determine a recalled resource from the plurality of resources based on the recall score of each resource.

[0148] This method fuses user behavior graphs, resource modal graphs, and knowledge graphs into a unified heterogeneous graph structure, breaking the limitations of traditional recommender systems that rely solely on user-resource behavior to construct graphs. By introducing a weighted path propagation mechanism, the system dynamically regulates structural paths from different sources (such as behavioral paths, modal paths, and knowledge paths), prioritizing information propagation along semantically consistent paths, effectively alleviating structural discontinuities and semantic disconnects. Furthermore, this method proposes a cold-start "graph replenishment" mechanism, allowing inactive users or new resource nodes to participate in propagation through modal features and knowledge entities, improving the system's recall coverage in cold-start and long-tail scenarios. Finally, a multi-source path fusion strategy aggregates propagated information from different graph types, enabling richer and more semantically consistent representation learning for nodes, significantly improving recommendation accuracy and system stability. This solution also exhibits strong modularity and scalability, allowing for flexible integration into existing graph recall systems.

[0149] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0150] It can be understood that the same / similar parts between the various embodiments of the above method in this specification can be referred to each other, and each embodiment focuses on the differences from other embodiments. For related parts, please refer to the description of other method embodiments.

[0151] Based on the same inventive concept, an embodiment of the present disclosure further provides a resource recall device for implementing the above-mentioned resource recall method.

[0152] Figure 8 FIG. 1 is a structural block diagram of a resource recall device according to an exemplary embodiment. Figure 8 , the device comprises:

[0153] A heterogeneous graph construction unit 810 is configured to construct a plurality of heterogeneous graphs based on a historical interaction dataset between a plurality of users and a plurality of resources; the plurality of heterogeneous graphs includes a behavior graph representing interaction records between users and resources, a modal graph representing relationships between resources under different modal features, and a knowledge graph representing relationships between resources under preset attribute information;

[0154] An information propagation unit 820 is configured to determine a target embedding representation for each user and a target embedding representation for each resource based on the connection relationship between nodes in each heterogeneous graph and the initial embedding representation of each node;

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

[0156] The resource screening unit 840 is configured to determine the recalled resources from the plurality of resources based on the recall score of each resource.

[0157] In an exemplary embodiment, each modal feature corresponds to a modal graph; the heterogeneous graph construction unit 810 is further configured to execute, for each modal feature, obtaining the similarity between each two resources in the multiple resources under the modal feature; taking each resource as a node, establishing a connecting edge between the nodes corresponding to two resources whose similarity is higher than a threshold, and obtaining 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 determine 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 a node, establishing connecting edges between associated resources and attribute information, and establishing connecting edges between associated 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 with a connection relationship 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 representation of each node, a multi-layer joint propagation operation of the 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, for any object among multiple users and multiple resources, determining the object embedding representation of the object in each heterogeneous graph in each layer; fusing the object embedding representations in each heterogeneous graph to obtain a single-layer embedding representation of the object in each layer; and performing weighted averaging on the single-layer embedding representations obtained in each layer to obtain a target embedding representation of the object.

[0161] In an exemplary embodiment, the information propagation unit 820 is further configured to execute, for the modal graph and knowledge graph in each heterogeneous graph, a preset propagation control function, and assign propagation weights between two nodes with a connection relationship in a manner such that the propagation weights are positively correlated with the semantic relevance.

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

[0163] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0164] Figure 9 1 is a block diagram of an electronic device 900 for implementing a resource recall method according to an exemplary embodiment. For example, the electronic device 900 may be a server. Figure 9 The electronic device 900 includes a processing component 920, which further includes one or more processors, and a memory resource represented by a memory 922 for storing instructions executable by the processing component 920, such as an application. The application stored in the memory 922 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 920 is configured to execute the instructions to perform the above method.

[0165] The electronic device 900 may further include a power supply component 924 configured to perform power management of the electronic device 900, a wired or wireless network interface 926 configured to connect the electronic device 900 to a network, and an input / output (I / O) interface 928. The electronic device 900 may operate based on an operating system stored in the memory 922, such as Windows 98, Mac OS X, Unix, Linux, FreeBSD, or the like.

[0166] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as memory 922 including instructions. The instructions may be executed by a processor of electronic device 900 to perform the above 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, optical data storage device, etc.

[0167] In an exemplary embodiment, a computer program product is further provided. The computer program product includes instructions, and the instructions can be executed by a processor of the electronic device 900 to implement the above method.

[0168] It should be noted that the above-mentioned devices, servers, computer-readable storage media, computer program products, etc. can also include other implementation methods according to the description of the method embodiments. The specific implementation methods can refer to the description of the relevant method embodiments and will not be described one by one here.

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

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

Claims

1. A resource recall method, characterized in that: include: Based on historical interaction datasets between multiple users and multiple resources, multiple heterogeneous graphs are constructed; the multiple heterogeneous graphs include a behavior graph representing the interaction records between users and resources, a modal graph representing the relationship between each resource under different modal characteristics, and a knowledge graph representing the relationship between each resource under preset attribute information; Based on the connection relationship between the 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; For each resource, calculate a recall score for the resource based on the target embedding representation of the resource and the target embedding representation of each user; A recalled resource is determined from the plurality of resources based on the recall score of each resource.

2. The method according to claim 1, characterized in that Each modal feature corresponds to a modal graph. Based on the historical interaction dataset, the process of constructing the modal graph includes: For each modal feature, obtaining the similarity between every two resources in the plurality of resources under the modal feature; Taking each resource as a node, a connecting edge is established between the nodes corresponding to two resources with a similarity higher than a threshold, and a modal graph under the modal feature is obtained.

3. The method according to claim 1, characterized in that There are multiple preset attribute information; based on the historical interaction data set, the process of constructing the knowledge graph includes: Determine the association between each resource and each attribute information, and the association between each attribute information; With each resource and each attribute information as a node, connecting edges are established between related resources and attribute information, and connecting edges are established between related attribute information to obtain the knowledge graph.

4. The method according to claim 1, wherein The determining of the target embedding representation of each user and the target embedding representation of each resource based on the connection relationship between the nodes in each heterogeneous graph and the initial embedding representation of each node includes: Using a preset propagation control function, a propagation weight is assigned between two nodes in a connection relationship in each heterogeneous graph; wherein, in the propagation control function, each heterogeneous graph corresponds to a different gating network; 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 a target embedding representation of each user and a target embedding representation of each resource.

5. The method according to claim 4, characterized in that The multi-layer joint propagation operation of the graph neural network is performed on each heterogeneous graph based on the propagation weight and the initial embedding representation of each node to obtain the target embedding representation of each user and the target embedding representation of each resource, including: For any object among the multiple users and the multiple resources, determining, in each layer, an object embedding representation of the object in each heterogeneous graph; The object embedding representations in each heterogeneous graph are fused to obtain the single-layer embedding representation of the object in each layer; The single-layer embedding representation obtained at each layer is weighted averaged to obtain the target embedding representation of the object.

6. The method according to claim 4, characterized in that The method of assigning a propagation weight between two nodes having a connection relationship in each heterogeneous graph by using a preset propagation control function includes: For the modal graphs and knowledge graphs in each heterogeneous graph, a preset propagation control function is used to assign propagation weights between two nodes with a connection relationship in a way that the propagation weights are positively correlated with the semantic relevance.

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

8. A resource recall device, characterized in that: include: A heterogeneous graph construction unit is configured to construct a plurality of heterogeneous graphs based on a historical interaction dataset between a plurality of users and a plurality of resources; The multiple heterogeneous graphs include a behavior 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; An information propagation unit is configured to determine a target embedding representation of each user and a target embedding representation of each resource based on the connection relationship between nodes in each heterogeneous graph and the initial embedding representation of each node; a score calculation unit configured to calculate, for each resource, a recall score for the resource based on a target embedding representation of the resource and a target embedding representation of each user; The resource screening unit is configured to determine the recalled resources from the plurality of resources based on the recall score of each resource.

9. A server, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the resource recall method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that When the instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the resource recall method according to any one of claims 1 to 7.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the resource recall method according to any one of claims 1 to 7 is implemented.

Images