Graph data processing method and device, computer readable storage medium and electronic equipment

Abstract

The embodiment of the invention discloses a graph data processing method and device, a computer readable storage medium and electronic equipment, and the method comprises the steps: employing a target deep learning model which comprises a graph neural network model, a hypergraph neural network model and a feature fusion model, processing first graph data of an incidence relation graph of the service entity and second graph data of an incidence relation hypergraph of the service entity so as to generate first feature data of the service entity according to the first graph data through a graph neural network model, and generating second feature data of the service entity according to the second graph data through a hypergraph neural network model, and performing feature fusion processing on the first feature data and the second feature data of the same service entity through a feature fusion model, thereby obtaining target feature data of the service entity.

Description

Technical Field

[0001] The present invention relates to the technical field of graph data processing, and in particular to a graph data processing method, device, computer-readable storage medium and electronic device. Background Art

[0002] A graph is a set of nodes and edges, where a "node" can represent an entity and an "edge" can represent the relationship between entities. Graph data is gaining more and more attention because it can simply and clearly express things and their interrelationships. For example, a normal graph can refer to a graph in which each edge connects 2 nodes. By analyzing a normal graph, we can reveal the connection pattern between nodes, explore the rules hidden behind the data, and help enterprises make more informed decisions. A hypergraph is an extension of a normal graph, which allows an edge to connect any number of nodes, so that the relationship between nodes can be described more flexibly. Since a hypergraph can provide more detailed and comprehensive information, it can be used in some complex business scenarios to help enterprises understand business data more deeply and discover hidden rules. Based on this, how to combine graph data of various types of graphs such as normal graphs and hypergraphs for collaborative analysis to improve the accuracy of the extracted representation features of business entities, and then facilitate the subsequent use of the representation features of business entities to assist enterprises in making business decisions, has become a technical problem that needs to be solved urgently. Summary of the invention

[0003] The embodiments of this specification provide a graph data processing method, device, computer-readable storage medium and electronic device, which can perform collaborative analysis on graph data of various types of graphs such as ordinary graphs and hypergraphs, which is beneficial to improving the accuracy of feature data extracted from business entities, and further facilitates enterprises to use accurate feature data of business entities to make accurate business decisions.

[0004] An embodiment of the present specification provides a method for processing graph data, the method comprising:

[0005] Acquire first graph data of an association relationship graph of a business entity; the nodes in the association relationship graph are the business entities, and the edges in the association relationship graph are used to reflect the first association relationship between two nodes connected by the edge;

[0006] Acquire second graph data of an association relationship hypergraph of the business entity; the nodes in the association relationship hypergraph are the business entities, and the hyperedges in the association relationship hypergraph are used to reflect the second association relationship between the nodes in the hyperedges;

[0007] Using a target deep learning model, processing the first graph data and the second graph data to obtain target feature data of the business entity;

[0008] The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model;

[0009] The graph neural network model is used to generate first feature data of the business entity based on the first graph data, the hypergraph neural network model is used to generate second feature data of the business entity based on the second graph data, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain the target feature data of the business entity.

[0010] A model training method provided in an embodiment of the present specification includes:

[0011] Obtaining a first graph data sample of an association relationship graph sample of a business entity sample; the nodes in the association relationship graph sample are the business entity samples, and the edges in the association relationship graph sample are used to reflect the first association relationship between two nodes connected by the edges;

[0012] Obtain a second graph data sample of an association relationship hypergraph sample of the business entity sample; the nodes in the association relationship hypergraph sample are the business entity samples, and the hyperedges in the association relationship hypergraph sample are used to reflect the second association relationship between the nodes in the hyperedges;

[0013] Using the first graph data sample and the second graph data sample to train an initial deep learning model to obtain a target deep learning model;

[0014] The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model;

[0015] The graph neural network model is used to generate first feature data of the business entity sample based on the first graph data sample, the hypergraph neural network model is used to generate second feature data of the business entity sample based on the second graph data sample, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity sample to obtain target feature data of the business entity sample.

[0016] The embodiment of this specification also provides a graph data processing device, including:

[0017] A first acquisition module is used to acquire first graph data of an association relationship graph of a business entity; the nodes in the association relationship graph are the business entities, and the edges in the association relationship graph are used to reflect the first association relationship between two nodes connected by the edge;

[0018] A second acquisition module is used to acquire second graph data of an association relationship hypergraph of the business entity; the nodes in the association relationship hypergraph are the business entities, and the hyperedges in the association relationship hypergraph are used to reflect the second association relationship between the nodes in the hyperedges;

[0019] A graph data processing module, configured to process the first graph data and the second graph data using a target deep learning model to obtain target feature data of the business entity;

[0020] The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model;

[0021] The graph neural network model is used to generate first feature data of the business entity based on the first graph data, the hypergraph neural network model is used to generate second feature data of the business entity based on the second graph data, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain the target feature data of the business entity.

[0022] The embodiment of this specification also provides a model training device, including:

[0023] A first acquisition module is used to acquire a first graph data sample of an association relationship graph sample of a business entity sample; the nodes in the association relationship graph sample are the business entity samples, and the edges in the association relationship graph sample are used to reflect the first association relationship between two nodes connected by the edges;

[0024] A second acquisition module is used to acquire a second graph data sample of an association relationship hypergraph sample of the business entity sample; the nodes in the association relationship hypergraph sample are the business entity samples, and the hyperedges in the association relationship hypergraph sample are used to reflect the second association relationship between the nodes in the hyperedges;

[0025] A training module, used to train an initial deep learning model using the first graph data sample and the second graph data sample to obtain a target deep learning model;

[0026] The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model;

[0027] The graph neural network model is used to generate first feature data of the business entity sample based on the first graph data sample, the hypergraph neural network model is used to generate second feature data of the business entity sample based on the second graph data sample, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity sample to obtain target feature data of the business entity sample.

[0028] The embodiments of this specification also provide a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0029] An embodiment of the present specification also provides an electronic device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the above method.

[0030] In an embodiment of the present specification, a target deep learning model including a graph neural network model, a hypergraph neural network model and a feature fusion model is used to process first graph data of a business entity's association relationship graph and second graph data of a business entity's association relationship hypergraph, so as to generate first feature data of the business entity based on the first graph data through the graph neural network model, generate second feature data of the business entity based on the second graph data through the hypergraph neural network model, and perform feature fusion processing on the first feature data and the second feature data of the same business entity through the feature fusion model to obtain target feature data of the business entity, thereby being able to perform collaborative analysis in combination with graph data such as ordinary graphs and hypergraphs that reveal the association relationships between business entities from different levels and angles, which is beneficial to improving the accuracy of the acquired target feature data of the business entity, thereby facilitating enterprises to make accurate business decisions using the accurate target feature data of business entities. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic diagram of an application scenario of a graph data processing solution provided in an embodiment of this specification;

[0032] Figure 2 A flowchart of a method for processing graph data provided in an embodiment of this specification;

[0033] Figure 3 A schematic diagram of the structure of a target deep learning model provided in an embodiment of this specification;

[0034] Figure 4 A flowchart of a model training method provided in an embodiment of this specification;

[0035] Figure 5 A schematic diagram of the structure of a graph data processing device provided in an embodiment of this specification;

[0036] Figure 6 A schematic diagram of the structure of a model training device provided in an embodiment of this specification;

[0037] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this specification. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0039] Before specifically describing the graph data processing solution provided in the embodiments of this specification, the relevant technical background is first described.

[0040] At present, service providers often generate a large amount of business data in the process of providing services to users. These business data can often reflect various types of relationships between business entities. When multiple relationships between business entities are depicted in a graph, it will appear messy and difficult to observe and learn. Therefore, layered graph technology can be used to depict various relationships between business entities. By taking business entities as nodes and the relationships between business entities as edges, and dividing nodes and edges according to different levels, a layered graph containing multiple levels is formed to clearly display the hierarchical structure and hierarchical relationship of business data or business entities.

[0041] For example, in the field of e-commerce, transportation or polymer, e-commerce participants, road segments or atoms contained therein can be used as business entities, and transaction relationships, road connection relationships or atom link relationships can be used as association relationships to construct the connection edges between two business entities to obtain the association relationship graph of business entities; and hyperedges between any number of business entities can be constructed to obtain the association relationship hypergraph of business entities. Thus, by combining the association relationship hypergraph of business entities with the association relationship graph, a comprehensive hierarchical graph data of business entities can be obtained.

[0042] However, currently available graph data processing solutions are usually unable to handle hierarchical graph data structures, and thus are unable to combine graph data of various types of graphs such as ordinary graphs and hypergraphs for collaborative analysis, making it impossible to extract accurate representation features of business entities, affecting subsequent business decisions of service providers.

[0043] See also Figure 1 , which is a schematic diagram of an application scenario of a graph data processing solution provided in an embodiment of this specification.

[0044] like Figure 1 As shown, a target deep learning model 102 including a graph neural network model, a hypergraph neural network model and a feature fusion model can be pre-deployed at device 101, so that after device 101 obtains the first graph data of the association relationship graph 103 of the business entity and the second graph data of the association relationship hypergraph 104 of the business entity, it can use the target deep learning model to process the first graph data and the second graph data, so as to generate the first feature data of the business entity according to the first graph data through the graph neural network model, generate the second feature data of the business entity according to the second graph data through the hypergraph neural network model, and perform feature fusion processing on the first feature data and the second feature data of the same business entity through the feature fusion model to obtain the target feature data of the business entity.

[0045] Among them, the nodes in the association graph 103 can be the business entities, and the edges in the association graph 103 can be used to reflect the first association relationship between the two nodes connected by the edges; and the nodes in the association hypergraph 104 are also the business entities, and the hyperedges in the association hypergraph can be used to reflect the second association relationship between the nodes in the hyperedges. By using the target deep learning model 102 to collaboratively analyze and process the graph data such as the association graph 103 and the association hypergraph 104 that reveal multiple association relationships between business entities from different levels and angles, it is beneficial to improve the accuracy of the target feature data of the acquired business entities, thereby facilitating enterprises to use the accurate target feature data of business entities to make accurate business decisions.

[0046] See also Figure 2 , is a flowchart of a graph data processing method provided in an embodiment of this specification. The execution subject of the process can be a device equipped with a target deep learning model, or a device with the call permission of the target deep learning model, or an application installed on the above device. Figure 2 The process shown is described in detail, and the graph data processing method may specifically include the following steps:

[0047] Step 202, obtaining first graph data of an association relationship graph of business entities; the nodes in the association relationship graph are business entities, and the edges in the association relationship graph are used to reflect the first association relationship between two nodes connected by the edge.

[0048] In the embodiments of this specification, a general graph (Graph) may refer to a graph in which the number of nodes connected by each edge is 2, and therefore, the association relationship graph of a business entity may belong to a general graph (Graph). Due to the types of business entities involved in different businesses, and the meanings of the first association relationships between business entities often differ, that is, the meanings of the nodes and edges in the association relationship graph of business entities obtained in different business scenarios can be set according to actual needs, and no specific limitation is made to this. In actual applications, the association relationship graph of a business entity can be either an undirected graph or a directed graph, and no specific limitation is made to this.

[0049] For ease of understanding, an example is given to illustrate this. In the field of e-commerce, business entities may include but are not limited to: buyer users, seller users, goods, devices used by users during transactions, payment accounts, delivery addresses, etc. The first association relationship between business entities may include but are not limited to: commodity purchase relationship, commodity selling relationship, capital expenditure relationship, capital inflow relationship, etc. In the field of transportation, business entities may include but are not limited to: road segments, intersections, traffic lights, roadside equipment, etc. The first association relationship between business entities may include but are not limited to: direct connection relationship of road segments, intersection relationship of equipment monitoring areas, equipment location relationship, etc. In the field of polymers, business entities may include: atoms, molecular monomers, etc. The first association relationship between business entities may include but are not limited to: covalent bond connection relationship, ionic bond connection relationship, metallic bond connection relationship, molecular bond connection relationship, and hydrogen bond connection relationship, etc.

[0050] In an embodiment of the present specification, the first graph data of the association relationship graph of business entities may generally include: initial feature data of each business entity contained in the association relationship graph, adjacency relationship data (for example, an adjacency matrix or adjacency list of a general graph) used to reflect the first association relationship between business entities, and the connecting edges between business entities can be determined based on the adjacency relationship data. In addition, preset feature data of the edges between business entities may also be included.

[0051] Among them, the types of initial feature data of the business entity and the preset feature data of the edges between business entities can be set according to actual needs; for example, when the business entity is a buyer / seller user, the initial feature data may include but is not limited to: user basic information, operation habit information, transaction frequency information; and the preset feature data of the edges between business entities may include: information reflecting the meaning and type of the first association relationship between business entities, such as commodity trading relationship, classmate relationship, social relationship, etc.; when the business entity is a road segment, the initial feature data may include but is not limited to: location information, width information, highway grade information, etc. of the road segment; and the preset feature data of the edges between business entities may include but are not limited to: entry relationship, exit relationship, etc., without specific limitation.

[0052] Step 204, obtaining second graph data of the association relationship hypergraph of the business entity; the nodes in the association relationship hypergraph are the business entities, and the hyperedges in the association relationship hypergraph are used to reflect the second association relationship between the nodes in the hyperedges.

[0053] In the embodiments of this specification, a hypergraph may be an extension of a common graph, and a hyperedge in a hypergraph may connect any number of nodes, so a hypergraph of business entity associations may belong to a hypergraph. Due to the types of business entities involved in different businesses, and the meanings of the second associations between business entities often differ, that is, the meanings of nodes and hyperedges in the hypergraph of business entity associations obtained in different business scenarios may be set according to actual needs, and no specific limitation is made to this.

[0054] For ease of understanding, an example of a hypergraph of business entity associations is given. For example, when the business entity is a buyer user, a hyperedge in the hypergraph may represent a set of buyer users who have used the same device.

[0055] Or, a set of buyer users who have used the same IP address, or a set of buyer users who have purchased the same product, or a set of buyer users who have browsed a specified product within a specified time period, etc.; when the business entity is a road segment, a hyperedge in the association hypergraph can represent a set of road segments belonging to a straight road, or a set of road segments that can form a connecting road between the first address and the second address, etc., and there is no specific limitation on this.

[0056] because Figure 2The middle solution needs to combine graph data of various types of graphs such as ordinary graphs and hypergraphs for collaborative analysis to improve the accuracy of the feature data of the extracted business entities. Therefore, the nodes (i.e., business entities) in the association relationship graph mentioned in step 202 and the association relationship hypergraph mentioned in step 204 should be consistent. For example, they can be the same user group, or the same batch of road segments, or the same set of atoms contained in the same polymer, etc. However, there are often differences in the association relationships between the business entities represented by the edges and hyperedges in the association relationship graph mentioned in step 202 and the association relationship hypergraph mentioned in step 204, which will not be elaborated on.

[0057] In the embodiments of this specification, the second graph data of the hypergraph of the association relationship of the business entities may generally include: initial feature data of each business entity included in the association relationship graph, association relationship data (e.g., an association matrix of the hypergraph) used to reflect the second association relationship between the business entities, and all business entities included in each hyperedge may be determined based on the association relationship data. In addition, it may also include preset feature data of the hyperedge, etc., which is not specifically limited.

[0058] Since the business entities involved in the association relationship graph in step 202 and the association relationship hypergraph in step 204 are consistent, the initial feature data of the business entities contained in the first graph data in step 202 and the second graph data in step 204 are usually consistent, which is not described in detail. The preset feature data of the hyperedge may include, but is not limited to: information reflecting the meaning and type of the second association relationship between the business entities, which is not described in detail.

[0059] Step 206: Use the target deep learning model to process the first graph data and the second graph data to obtain target feature data of the business entity.

[0060] Among them, the target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model.

[0061] The graph neural network model is used to generate first feature data of the business entity based on the first graph data, the hypergraph neural network model is used to generate second feature data of the business entity based on the second graph data, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain the target feature data of the business entity.

[0062] In the embodiments of this specification, the graph neural network model (Graph Neural Network, GNN) can be a general term for algorithms that use neural networks to learn graph structure data, extract and explore features and patterns in graph structure data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. In practical applications, the graph neural network model may include but is not limited to: Graph Convolutional Network (Graph Convolutional Network, GCN), Graph Attention Network (Graph Attention Network, GAT), Graph Isomorphism Network (Graph Isomorphism Network, GIN), GraphSage network, principal neighborhood aggregation network (Principal Neighbourhood Aggregation, PNA), etc. Since the graph neural network model is usually applicable to graph data in ordinary graphs, the representation of nodes is updated by aggregating the neighbor information of nodes, thereby learning the structure and spatial features of ordinary graphs. Based on this, the graph neural network model can be used to process the first graph data of the association relationship graph of the business entity obtained in step 202 to generate the first feature data of the business entity.

[0063] In the embodiments of this specification, the hypergraph neural network model (Hypergraph Neural Nerworks, HGNN) can be regarded as an extension of the graph neural network, which can process hypergraphs of different sizes and shapes, and can embed and represent nodes and hyperedges in the hypergraph. Compared with traditional graph neural networks, hypergraph neural networks can handle multivariate relationships and high-order relationships, and show better performance in some applications. Based on this, the hypergraph neural network model can be used to process the second graph data of the association relationship hypergraph of the business entity obtained in step 204 to generate the second feature data of the business entity.

[0064] In the embodiments of the present specification, since the business entity to which the first feature data extracted by the graph neural network model belongs is consistent with the business entity to which the second feature data extracted by the hypergraph neural network model belongs, that is, for the same business entity, it has two feature data, the first feature data and the second feature data. Therefore, the feature fusion model can also be used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain target feature data with good accuracy for the business entity.

[0065] The feature fusion model can be implemented using a single fully connected layer, or a multi-layer perceptron (MLP), or a neural network model with a more complex structure, without specific limitation. In practical applications, the first feature data and the second feature data can be concatenated using the concat method to obtain the target feature data, or other methods can be used to achieve the fusion of the first feature data and the second feature data to obtain the target feature data, without specific limitation.

[0066] In the embodiments of the present specification, a target deep learning model can be built and trained in advance using a graph neural network model, a hypergraph neural network model and a feature fusion model, so that when executing step 206, the target deep learning model can be called to process the first graph data and the second graph data, thereby generating target feature data of the business entity.

[0067] Figure 2 In the method, since hypergraphs can better handle complex relationships between nodes, while ordinary graphs are more suitable for handling simple node associations, by combining ordinary graphs and hypergraphs to reveal the association relationships between business entities from different levels and angles for collaborative analysis, it can not only analyze the relevant business data of business entities more comprehensively and deeply, making the generated target feature data of business realization more comprehensive and accurate, but also can flexibly respond to the complexity of data in different scenarios, thereby helping enterprises improve decision-making accuracy and efficiency, and has good practicality.

[0068] based on Figure 2 The method in this specification also provides some specific implementation plans of the method, which are described below.

[0069] For ease of understanding, the specific implementation method of the target deep learning model is explained here.

[0070] Optionally, the graph neural network model can be a model for processing graph data based on an efficient Transformer architecture. The input of the graph neural network model can be the initial feature data of the business entity in the first graph data and the adjacency relationship data between the business entities, and the output of the graph neural network model can be the first feature data of the business entity.

[0071] Optionally, the hypergraph neural network model may include at least one hypergraph neural network module built on an efficient Transformer architecture for processing hypergraph data, and a feature fusion module. The hypergraph neural network module is connected to the feature fusion module. For any one of the hypergraph neural network modules, the input of the hypergraph neural network module may be the initial feature data of each of the business entities contained in a hyperedge in the second graph data, and the output of the hypergraph neural network module may be the third feature data of the business entity contained in the hyperedge. After the feature fusion sub-model performs feature fusion processing on the received third feature data of the same business entity, it can output the second feature data of the business entity.

[0072] Among them, the computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture.

[0073] In the embodiments of this specification, the core feature of the Transformer architecture may be the use of a Self-Attention structure to replace the RNN network structure commonly used in NLP tasks, so that not only can the meaning of the input sequence be understood more accurately, but also parallel calculations can be performed, thereby improving computational efficiency. However, since the self-attention mechanism of the Transformer architecture has quadratic time and space complexity when processing long sequences, this means that as the length of the input sequence increases, the required computing resources and time increase exponentially. At the same time, due to the large number of nodes and edges in a large-scale graph, which may contain tens of thousands or even more nodes and edges, and in actual applications, the association graph and association hypergraph of business entities may belong to large-scale graphs, it may not be possible to use the graph neural network model and hypergraph neural network model built on the Transformer architecture to build the target deep learning model.

[0074] In order to solve the problem of high computational complexity of the Transformer architecture, the efficient Transformer architecture came into being. The efficient Transformer architecture can refer to a model architecture whose computational complexity (for example, time complexity, space complexity) is lower than that of the Transformer architecture by optimizing the algorithm and improving the network structure of the Transformer architecture. In practical applications, the efficient Transformer architecture can include but is not limited to any one of: Linear Transformer, Fast Transformer, Reformer, Linformer, Performer.

[0075] By using the efficient Transformer architecture to build the graph neural network model and the hypergraph neural network model in the target deep learning model, the computational complexity of the model can be greatly reduced, so that the graph neural network model and the hypergraph neural network model can support graph-scale learning calculations of millions or even more nodes. In this way, the target deep learning model can be used to efficiently generate target feature data for each business entity, thereby meeting the real-time requirements of business operations.

[0076] It can be understood that, since each hyperedge in the hypergraph contains at least part of the nodes in the hypergraph, a hyperedge can be regarded as a node subset of the hypergraph, that is, a hyperedge can be regarded as a normal graph in which the nodes contained are connected in pairs. Based on this, a corresponding hypergraph neural network module can be set in the hypergraph neural network model for a hyperedge in the hypergraph of the association relationship of the business entity, so that the hypergraph neural network module can analyze and process the initial feature data of each business entity in the hyperedge to extract the third feature data of each business entity. In addition, since different hyperedges may contain the same business entity, it is possible that multiple hypergraph neural network modules have extracted the third feature data of the same business entity. Based on this, it is also necessary to set a feature fusion module in the hypergraph neural network model so that the feature fusion module performs feature fusion processing on the third feature data of the same business entity output by each hypergraph neural network module to obtain the second feature data of this business entity, which is conducive to improving the computational efficiency of the hypergraph neural network model. Among them, the types of hyperedges processed by different hypergraph neural network modules can be different; and the feature fusion module can be implemented by a single fully connected layer or a multi-layer perceptron MLP, and there is no specific limitation on this.

[0077] In practical applications, the graph neural network model may include multiple first network layers that are connected in sequence and built based on an efficient Transformer architecture. Among them, after the first first network layer receives the initial feature data and the adjacency relationship data of the business entity, it can perform feature update processing on the business entity and output the first updated feature data of the business entity, and the output of the previous first network layer and the adjacency relationship data can be used as the input of the next first network layer, and the output of the last first network layer can be the first feature data of the business entity.

[0078] Figure 3 This is a schematic diagram of the structure of a target deep learning model provided in the embodiments of this specification. Figure 3As shown, the target deep learning model may include a graph neural network model 301. The structures and working principles of the multiple first network layers connected in series in the graph neural network model 301 may be consistent, and any first network layer may aggregate the feature information of the neighboring nodes of a node to update the representation of the node, thereby learning the structure and spatial features of the common graph. Therefore, the multiple first network layers connected in series in the graph neural network model 301 may continuously iteratively update the feature data of the business entity, and use the output of the last first network layer as the required input to the first feature data of the business entity of the feature fusion model 303.

[0079] In practical applications, the hypergraph neural network module may include multiple second network layers connected in sequence and built based on an efficient Transformer architecture. After the first second network layer receives the initial feature data of each business entity contained in a hyperedge, it can perform feature update processing on the business entity and output the second updated feature data of the business entity, and the output of the previous second network layer can be used as the input of the next second network layer, and the output of the last second network layer can be the third feature data of the business entity.

[0080] like Figure 3 As shown, the target deep learning model may include a hypergraph neural network model 302. The hypergraph neural network model 302 may include multiple parallel hypergraph neural network modules, wherein the structures and working principles of the multiple second network layers connected in series in each hypergraph neural network module may be consistent, and any second network layer may aggregate the feature information of the neighboring nodes of a node to update the representation of the node, thereby learning the structure and spatial features of the hypergraph. Therefore, the multiple second network layers connected in series in the hypergraph neural network module may continuously iterate and update the feature data of the business entity, and use the output of the last second network layer as the third feature data of the business entity required to be input to the feature fusion module, so that the feature fusion module may input the obtained second feature data of the business entity into the feature fusion model 303 after fusing the third feature data of the same business entity.

[0081] Optionally, the first network layer can be specifically used to: for any one of the business entities, use the Gumbel-Softmax algorithm to filter out a first neighbor node from the neighbor nodes of the business entity, so as to perform feature update processing on the business entity according to feature data of the first neighbor node.

[0082] Optionally, the second network layer can be specifically used to: for any one of the business entities contained in a hyperedge, filter out a second neighbor node from the neighbor nodes of the business entity using the Gumbel-Softmax algorithm, so as to perform feature update processing on the business entity based on feature data of the second neighbor node.

[0083] In the embodiments of this specification, the Gumbel-Softmax algorithm is a method for extracting samples from a category with a specific probability. Specifically, when the first network layer and the second network layer need to use the feature data of the neighboring nodes to update the features of a certain business entity, the Gumbel-Softmax algorithm can be used to calculate the sampling probability of each neighboring node, and the preset number of neighboring nodes with the highest sampling probability are used as the first neighboring node / second neighboring node, so as to use the feature data of the first neighboring node / second neighboring node to update the feature data of the business entity. Since the first neighboring node / second neighboring node selected based on the Gumbel-Softmax algorithm has a certain randomness, and the first neighboring node / second neighboring node is usually closely related to the business entity, the accuracy of the updated feature data of the generated business entity can be guaranteed. In addition, compared with the method of updating the features of the business entity according to the feature data of all neighboring nodes, its operating efficiency is also higher, which can better meet the real-time requirements of business operation.

[0084] In an embodiment of the present specification, the target deep learning model may include: multiple feature extraction models connected in sequence; wherein, for any one of the feature extraction models, the feature fusion model contained in the feature extraction model can be connected to one of the graph neural network models and one of the hypergraph neural network models respectively.

[0085] After the first feature extraction model receives the first graph data and the second graph data, it performs feature update processing on the business entity and can output the third updated feature data of the business entity. In addition, the output of the previous feature extraction model can be used as the input of the next feature extraction model, and the output of the last feature extraction model can be the target feature data of the business entity.

[0086] In the embodiments of this specification, a feature fusion model can be used to connect a graph neural network model and a hypergraph neural network model respectively to serve as a feature extraction model. Since a feature extraction model can fuse the feature information of the business entity sample's association relationship graph and the business entity contained in the association relationship hypergraph, based on this, in order to better fuse the feature information of the business entity contained in the association relationship graph and the association relationship hypergraph, multiple sequentially connected feature extraction models can also be deployed in the target deep learning model.

[0087] like Figure 3 As shown, the target deep learning model may include multiple feature extraction models 31 with consistent structures and working principles. Any feature extraction model 31 can update and fuse the feature information of the business entity contained in the association relationship graph of the business entity sample and the association relationship hypergraph. Therefore, the multiple feature extraction models 31 connected in series in the target deep learning model can continuously iteratively update the feature data of the business entity, and use the output of the last feature extraction model 31 as the target feature data of the business entity.

[0088] It is worth noting that the inputs of the graph neural network model and the hypergraph neural network model in each feature extraction model contain consistent feature data of business entities. In addition, the inputs of the graph neural network model in each feature extraction model can also contain consistent adjacency relationship data between business entities, which will not be elaborated on here.

[0089] In practical applications, there is often a need to use the target feature data of the business entity output by the feature fusion model in the feature extraction model 31 to perform business analysis. Based on this, the target deep learning model can also include: a target classification model; the target classification model can be connected to the feature fusion model (in the feature extraction model 31).

[0090] Correspondingly, step 206: using the target deep learning model to process the first graph data and the second graph data may also include: using the target classification model to process the target feature data of one or more of the business entities output by the feature fusion model to obtain business processing results of preset categories related to the one or more business entities.

[0091] In the embodiments of this specification, the business processing results of the preset category may include but are not limited to: the classification results of whether the business entity belongs to the preset category, and at least one of the identification results of whether there is an association relationship of the preset category between multiple business entities. Since the meanings of the business processing results of the preset categories required to be generated for different business scenarios are often different, no specific limitation is made to this. For ease of understanding, examples are given to illustrate this. For example, the classification result of whether the buyer user business entity belongs to a fraudulent user, the classification result of whether the designated commodity business entity belongs to a commodity with a sales growth trend, the classification result of whether the target road section belongs to a busy traffic section, etc. Or, the identification result of whether the first user and the second user and other business entities have a classmate relationship, the identification result of whether the first device and the second device and other business entities belong to the same user, etc. There is no specific limitation on this.

[0092] In practical applications, since the data dimension of the input features of the target classification model may be inconsistent with the data dimension of the output features of the feature fusion model (in the feature extraction model 31), the target deep learning model may further include: a target decoder. Figure 3 As shown, the target decoder 32 can be connected to the feature fusion model (in the last feature extraction model 31) and the target classification model 33 respectively.

[0093] After receiving the target feature data of the business entity output by the feature fusion model, the target decoder 32 may output feature data of preset dimensions of one or more of the business entities to the target classification model.

[0094] The target classification model 33 can be specifically used to process the feature data of the preset dimensions of the one or more business entities to obtain business processing results of preset categories related to the one or more business entities.

[0095] In the embodiment of this specification, the target decoder 32 can be implemented by a fully connected layer or a multi-layer perceptron, which will not be described in detail. Also, the above-mentioned preset dimensions can be set according to actual needs, and there is no specific limitation on this.

[0096] Based on the same idea, see Figure 4 , is a flow chart of a model training method provided in an embodiment of this specification. The execution subject of the process can be a model training device, or an application installed on the model training device. Figure 4 The model training method may include the following steps:

[0097] Step 402, obtaining a first graph data sample of an association relationship graph sample of a business entity sample; the nodes in the association relationship graph sample are the business entity samples, and the edges in the association relationship graph sample are used to reflect the first association relationship between the two nodes connected by the edges.

[0098] In this embodiment of the specification, the business entity sample in step 402 is Figure 2 The type of business entity in step 202 may be consistent, the type of the association graph sample in step 402 may be consistent with the type of the association graph in step 202, and the type of the first graph data sample in step 402 may be consistent with the type of the first graph data in step 202, which will not be elaborated on.

[0099] Step 404, obtaining a second graph data sample of an association relationship hypergraph sample of the business entity sample; the nodes in the association relationship hypergraph sample are the business entity samples, and the hyperedges in the association relationship hypergraph sample are used to reflect the second association relationship between the nodes in the hyperedges.

[0100] In the embodiment of this specification, the business entity sample in step 404 is Figure 2 The types of business entities in step 204 may be consistent, the types of the association relationship hypergraph samples in step 404 and the association relationship hypergraph in step 204 may also be consistent, and the types of the second graph data samples in step 404 and the second graph data in step 204 may also be consistent, which will not be elaborated on.

[0101] Step 406: Use the first graph data sample and the second graph data sample to train the initial deep learning model to obtain a target deep learning model.

[0102] The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model;

[0103] The graph neural network model is used to generate first feature data of the business entity sample based on the first graph data sample, the hypergraph neural network model is used to generate second feature data of the business entity sample based on the second graph data sample, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity sample to obtain target feature data of the business entity sample.

[0104] In the embodiment of this specification, the target deep learning model trained in step 406 may be the target deep learning model used in step 206, which will not be elaborated on.

[0105] Figure 4 The method in the invention uses the first graph data sample of the association graph sample of the business entity sample and the second graph data sample of the association hypergraph sample to pre-train the initial deep learning model to obtain the target deep learning model, which is beneficial to improve the performance of the target deep learning model, and then ensure the implementation based on the target deep learning model. Figure 2 The accuracy and effectiveness of graph data processing solutions in .

[0106] In the actual business operation process, it is necessary not only to use the target deep learning model to extract the target feature data of the business entity, but also to perform business analysis based on the target feature data to obtain the business processing results of the preset category required for the business operation process. Therefore, the target deep learning model can also include a target classification model.

[0107] In the embodiments of this specification, an end-to-end approach can be used to train a target deep learning model including a target classification model. Specifically, the target deep learning model can also include: a target classification model; the feature fusion model can be connected to the target classification model; Figure 4 The method may also include:

[0108] Determine the label data that the first graph data sample and the second graph data sample need to carry; the label data is used to reflect the benchmark business processing results of a preset category related to one or more of the business entity samples.

[0109] Correspondingly, step 406: training the initial deep learning model using the first graph data sample and the second graph data sample to obtain a target deep learning model may include:

[0110] The initial deep learning model is trained using the first graph data samples and the second graph data samples carrying the label data to obtain predicted business processing results of preset categories related to the one or more business entity samples; wherein the predicted business processing results are obtained by the target classification model performing classification processing on the target feature data of the one or more business entity samples output by the feature fusion model.

[0111] Taking minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal, the parameters of the initial deep learning model are adjusted to obtain the target deep learning model.

[0112] In the embodiments of this specification, the structures of the initial deep learning model and the target deep learning model can be consistent, that is, the initial graph neural network model and the initial hypergraph neural network model contained in the initial deep learning model can be respectively connected to the initial feature fusion model, and the initial feature fusion model can be connected to the initial classification model. By inputting the first graph data sample and the second graph data sample into the initial graph neural network model and the initial hypergraph neural network model in the initial deep learning model respectively, the predicted business processing results of preset categories related to one or more business entity samples output by the initial classification model can be obtained. Therefore, based on the difference between the predicted business processing results and the preset benchmark business processing results, the parameters of the entire initial deep learning model can be optimized to obtain the target deep learning model. Convenient and fast.

[0113] In practical applications, since the data dimension of the input features of the target classification model may be inconsistent with the data dimension of the output features of the feature fusion model, the target deep learning model may also include: a target decoder; the target decoder may be connected to the feature fusion model and the target classification model respectively.

[0114] After receiving the target feature data of the business entity samples output by the feature fusion model, the target decoder may output the feature data of the preset dimension of the one or more business entity samples to the target classification model.

[0115] The target classification model can be specifically used to: perform classification processing on the feature data of preset dimensions of the one or more business entity samples to obtain the predicted business processing result.

[0116] It is understandable that the initial deep learning model can also include an initial decoder, and respectively connect the initial feature fusion model and the initial classification model to achieve the unification of feature dimensions. Moreover, when the initial decoder is included in the initial deep learning model, the end-to-end model training method can still be used. When the parameters of the entire initial deep learning model are optimized based on the difference between the predicted business processing results generated by the initial deep learning model and the benchmark business processing results, the target deep learning model can be directly trained to obtain the target deep learning model. This is convenient and quick, and no further explanation is given.

[0117] In practical applications, in order to improve the data fusion effect between ordinary graphs and hypergraphs, the model obtained by connecting a graph neural network model and a hypergraph neural network model to the same feature fusion model can be called a feature extraction model, and the initial deep learning model can contain multiple feature extraction models connected in sequence, which will not be elaborated here.

[0118] In the embodiments of this specification, a phased model training method can also be adopted, first training a feature extraction model for extracting target feature data of business entity samples, and then training a target classification model separately, and then splicing the feature extraction model and the target classification model to obtain a target deep learning model.

[0119] Based on this, the target deep learning model may include: a feature extraction model; the feature fusion model contained in the feature extraction model may be connected to one of the graph neural network models and one of the hypergraph neural network models respectively. The initial deep learning model may be an autoencoder including an encoder and a decoder, and the encoder is the feature extraction model to be trained.

[0120] Correspondingly, step 406: training the initial deep learning model using the first graph data sample and the second graph data sample to obtain a target deep learning model may include:

[0121] The initial deep learning model is trained using the first graph data sample and the second graph data sample to obtain reconstructed feature data of the business entity sample output by the initial deep learning model.

[0122] Taking minimizing the difference between the initial feature data and the reconstructed feature data of the business entity samples in the first graph data sample and the second graph data sample as the training goal, the parameters of the initial deep learning model are adjusted to obtain the trained initial deep learning model.

[0123] Extract the encoder in the trained initial deep learning model to obtain the feature extraction model in the target deep learning model.

[0124] In the embodiments of this specification, an autoencoder (AE) is a type of artificial neural network used in semi-supervised learning and unsupervised learning, which uses the input information as a learning target to perform characterization learning on the input information. Based on this, the feature extraction model to be trained can be used as the encoder of the autoencoder, and the decoder of the autoencoder can be a multi-layer perceptron, so that the autoencoder is unsupervisedly trained based on the first image data sample and the second image data sample to obtain a trained feature extraction model.

[0125] Optionally, the target deep learning model may further include: a target classification model; Figure 4 The method may also include:

[0126] The first graph data sample and the second graph data sample are processed using the feature extraction model in the target deep learning model to obtain feature data samples of the business entity samples.

[0127] Determine label data carried by characteristic data samples of one or more of the business entity samples; the label data is used to reflect benchmark business processing results of a preset category related to the one or more business entity samples.

[0128] The initial classification model is trained using the feature data samples of the one or more business entity samples carrying the label data to obtain predicted business processing results of preset categories related to the one or more business entity samples.

[0129] Taking minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal, the parameters of the initial classification model are adjusted to obtain the target classification model.

[0130] The target deep learning model is generated according to the feature extraction model and the target classification model.

[0131] Similarly, the target deep learning model may further include: a target decoder; the target decoder may be connected to the target classification model.

[0132] The step of training the initial classification model by using the characteristic data samples of the one or more business entity samples carrying the label data to obtain the predicted business processing results of the preset categories related to the one or more business entity samples may specifically include:

[0133] The model obtained by connecting the initial decoder and the initial classification model is trained using the characteristic data samples of the one or more business entity samples carrying the label data to obtain the predicted business processing results of the preset category related to the one or more business entity samples.

[0134] The step of adjusting the parameters of the initial classification model with minimizing the difference between the predicted service processing result and the benchmark service processing result as the training goal to obtain the target classification model may specifically include:

[0135] Taking minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal, the parameters of the model obtained by connecting the initial decoder and the initial classification model are adjusted to obtain the target decoder and the target classification model.

[0136] Generating a target deep learning model according to the feature extraction model and the target classification model may specifically include:

[0137] Generate a target deep learning model based on the feature extraction model, target decoder and target classification model.

[0138] In the embodiments of the present specification, a feature extraction model is obtained by training in the first model training stage, and a target classification model or a model obtained by connecting a target decoder and a target classification model is obtained by training in the second model training stage, so that a target deep learning model can be generated by connecting the models obtained in the above two training stages, which has good flexibility.

[0139] In practical applications, since the target deep learning model may contain multiple sequentially connected feature extraction models, the encoder in the autoencoder used in the first model training stage can be composed of multiple sequentially connected feature extraction models to be trained, which will not be elaborated here.

[0140] Optionally, the graph neural network model can be a model for processing graph data based on an efficient Transformer architecture. The input of the graph neural network model can be the initial feature data of the business entity sample in the first graph data sample and the adjacency relationship data between the business entity sample, and the output of the graph neural network model can be the first feature data of the business entity sample.

[0141] Optionally, the hypergraph neural network model may include at least one hypergraph neural network module built on an efficient Transformer architecture for processing hypergraph data, and a feature fusion module; and the hypergraph neural network module may be connected to the feature fusion module.

[0142] For any one of the hypergraph neural network modules, the input of the hypergraph neural network module can be the initial feature data of each of the business entity samples contained in a hyperedge in the second graph data sample, and the output of the hypergraph neural network module can be the third feature data of the business entity sample contained in the hyperedge.

[0143] After the feature fusion sub-model performs feature fusion processing on the received third feature data of the same business entity sample, the second feature data of the business entity sample can be output.

[0144] Optionally, the graph neural network model may include a plurality of first network layers that are connected in sequence and are built based on the efficient Transformer architecture. After the first first network layer receives the initial feature data and the adjacency relationship data of the business entity sample, it may perform feature update processing on the business entity sample and output the first updated feature data of the business entity sample, and the output of the previous first network layer and the adjacency relationship data may be used as the input of the next first network layer, and the output of the last first network layer may be the first feature data of the business entity sample.

[0145] Optionally, the hypergraph neural network module may include multiple second network layers that are connected in sequence and built based on the efficient Transformer architecture. After the first second network layer receives the initial feature data of each business entity sample contained in a hyperedge, it can perform feature update processing on the business entity sample and output the second updated feature data of the business entity sample, and the output of the previous second network layer can be used as the input of the next second network layer, and the output of the last second network layer can be the third feature data of the business entity sample.

[0146] Among them, the efficient Transformer architecture may include: any one of FastTransformer, LinearTransformer, Performer, Linformer, and Reformer.

[0147] Optionally, the first network layer can be specifically used to: for any one of the business entity samples, use the Gumbel-Softmax algorithm to filter out a first neighbor node from the neighbor nodes of the business entity sample, so as to perform feature update processing on the business entity sample according to feature data of the first neighbor node.

[0148] Optionally, the second network layer can be specifically used to: for any one of the business entity samples contained in a hyperedge, filter out a second neighbor node from the neighbor nodes of the business entity sample using the Gumbel-Softmax algorithm, so as to perform feature update processing on the business entity sample based on feature data of the second neighbor node.

[0149] Based on the same idea, see Figure 5 , is a schematic diagram of the structure of a graph data processing device provided in an embodiment of this specification. Figure 5As shown, the graph data processing device 5 can be implemented as all or part of the electronic device through software, hardware or a combination of both. According to some embodiments, the graph data processing device 5 may include a first acquisition module 51, a second acquisition module 52, and a graph data processing module 53, wherein:

[0150] The first acquisition module 51 is used to acquire first graph data of an association relationship graph of a business entity; the nodes in the association relationship graph are the business entities, and the edges in the association relationship graph are used to reflect the first association relationship between the two nodes connected by the edge.

[0151] The second acquisition module 52 is used to obtain second graph data of the association relationship hypergraph of the business entity; the nodes in the association relationship hypergraph are the business entities, and the hyperedges in the association relationship hypergraph are used to reflect the second association relationship between the nodes in the hyperedges.

[0152] The graph data processing module 53 is used to process the first graph data and the second graph data using a target deep learning model to obtain target feature data of the business entity.

[0153] Among them, the target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model.

[0154] The graph neural network model is used to generate first feature data of the business entity based on the first graph data, the hypergraph neural network model is used to generate second feature data of the business entity based on the second graph data, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain the target feature data of the business entity.

[0155] Optionally, the graph neural network model is a model built based on the efficient Transformer architecture for processing graph data. The input of the graph neural network model is the initial feature data of the business entity in the first graph data and the adjacency relationship data between the business entities, and the output of the graph neural network model is the first feature data of the business entity.

[0156] Optionally, the graph neural network model includes multiple first network layers connected in sequence and built based on the efficient Transformer architecture; wherein, after the first first network layer receives the initial feature data and the adjacency relationship data of the business entity, it performs feature update processing on the business entity and outputs the first updated feature data of the business entity, and the output of the previous first network layer and the adjacency relationship data are used as inputs for the next first network layer, and the output of the last first network layer is the first feature data of the business entity.

[0157] Among them, the computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture. The efficient Transformer architecture may include any one of: Linear Transformer, Fast Transformer, Performer, Linformer, and Reformer.

[0158] Optionally, the first network layer can be specifically used to: for any one of the business entities, use the Gumbel-Softmax algorithm to filter out a first neighbor node from the neighbor nodes of the business entity, so as to perform feature update processing on the business entity according to feature data of the first neighbor node.

[0159] Optionally, the hypergraph neural network model includes at least one hypergraph neural network module built on an efficient Transformer architecture for processing hypergraph data, and a feature fusion module; wherein the hypergraph neural network module is connected to the feature fusion module; for any one of the hypergraph neural network modules, the input of the hypergraph neural network module is the initial feature data of each of the business entities contained in a hyperedge in the second graph data, and the output of the hypergraph neural network module is the third feature data of the business entity contained in the hyperedge; the feature fusion sub-model performs feature fusion processing on the received third feature data of the same business entity, and outputs the second feature data of the business entity; the computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture.

[0160] Optionally, the hypergraph neural network module includes multiple second network layers connected in sequence and built based on the efficient Transformer architecture; wherein, after the first second network layer receives the initial feature data of each business entity contained in a hyperedge, it performs feature update processing on the business entity and outputs the second updated feature data of the business entity, and the output of the previous second network layer serves as the input of the next second network layer, and the output of the last second network layer is the third feature data of the business entity.

[0161] Optionally, the second network layer can be specifically used to: for any one of the business entities contained in a hyperedge, filter out a second neighbor node from the neighbor nodes of the business entity using the Gumbel-Softmax algorithm, so as to perform feature update processing on the business entity based on feature data of the second neighbor node.

[0162] Optionally, the target deep learning model may include: multiple feature extraction models connected in sequence; wherein, for any one of the feature extraction models, the feature fusion model contained in the feature extraction model is respectively connected to one of the graph neural network models and one of the hypergraph neural network models; after the first feature extraction model receives the first graph data and the second graph data, it performs feature update processing on the business entity and outputs the third updated feature data of the business entity, and the output of the previous feature extraction model is used as the input of the next feature extraction model, and the output of the last feature extraction model is the target feature data of the business entity.

[0163] Optionally, the target deep learning model may also include: a target classification model; the feature fusion model is connected to the target classification model. Correspondingly, the graph data processing module may be specifically used to: utilize the target classification model to process the target feature data of one or more of the business entities output by the feature fusion model, and obtain business processing results of preset categories related to the one or more business entities. Among them, the business processing results of preset categories may include: a classification result for whether a business entity belongs to a preset category, and at least one of the identification results for whether multiple business entities have an association relationship of preset categories.

[0164] Optionally, the target deep learning model may also include: a target decoder; the target decoder is respectively connected to the feature fusion model and the target classification model; after the target decoder receives the target feature data of the business entity output by the feature fusion model, it outputs the feature data of preset dimensions of one or more of the business entities to the target classification model; the target classification model is specifically used to: process the feature data of preset dimensions of the one or more business entities to obtain business processing results of preset categories related to the one or more business entities.

[0165] Based on the same idea, see Figure 6 , is a schematic diagram of the structure of a model training device provided in an embodiment of this specification. Figure 6 As shown, the model training device 6 can be implemented as all or part of the electronic device through software, hardware or a combination of both. According to some embodiments, the model training device 6 may include a first acquisition module 61, a second acquisition module 62, and a training module 63, wherein:

[0166] The first acquisition module 61 is used to obtain a first graph data sample of an association relationship graph sample of a business entity sample; the nodes in the association relationship graph sample are the business entity samples, and the edges in the association relationship graph sample are used to reflect the first association relationship between the two nodes connected by the edge.

[0167] The second acquisition module 62 is used to obtain a second graph data sample of the association relationship hypergraph sample of the business entity sample; the nodes in the association relationship hypergraph sample are the business entity samples, and the hyperedges in the association relationship hypergraph sample are used to reflect the second association relationship between the nodes in the hyperedges.

[0168] The training module 63 is used to train the initial deep learning model using the first graph data sample and the second graph data sample to obtain a target deep learning model.

[0169] Among them, the target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model.

[0170] The graph neural network model is used to generate first feature data of the business entity sample based on the first graph data sample, the hypergraph neural network model is used to generate second feature data of the business entity sample based on the second graph data sample, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity sample to obtain target feature data of the business entity sample.

[0171] Optionally, the target deep learning model may also include: a target classification model; the feature fusion model is connected to the target classification model. Figure 6 The device may also include:

[0172] The first label determination module is used to determine the label data that the first graph data sample and the second graph data sample need to carry; the label data is used to reflect the benchmark business processing results of a preset category related to one or more business entity samples.

[0173] Training modules may include:

[0174] The first training unit is used to train the initial deep learning model using the first graph data samples and the second graph data samples carrying the label data to obtain predicted business processing results of preset categories related to one or more business entity samples; wherein the predicted business processing results are obtained by classifying the target feature data of the one or more business entity samples output by the feature fusion model by the target classification model.

[0175] The first optimization unit is used to adjust the parameters of the initial deep learning model with minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal to obtain the target deep learning model.

[0176] Optionally, the target deep learning model may further include: a target decoder; the target decoder is connected to the feature fusion model and the target classification model respectively. After receiving the target feature data of the business entity sample output by the feature fusion model, the target decoder outputs the feature data of the preset dimensions of the one or more business entity samples to the target classification model; the target classification model is specifically used to: perform classification processing on the feature data of the preset dimensions of the one or more business entity samples to obtain the predicted business processing result.

[0177] Optionally, the target deep learning model may include: a feature extraction model; the feature fusion model contained in the feature extraction model is respectively connected to a graph neural network model and a hypergraph neural network model; the initial deep learning model is an autoencoder including an encoder and a decoder, and the encoder is the feature extraction model to be trained.

[0178] Optionally, the training module may include:

[0179] The second training unit is used to train the initial deep learning model using the first graph data samples and the second graph data samples to obtain reconstructed feature data of the business entity samples output by the initial deep learning model.

[0180] The second optimization unit is used to adjust the parameters of the initial deep learning model with the training goal of minimizing the difference between the initial feature data and the reconstructed feature data of the business entity samples in the first graph data samples and the second graph data samples to obtain the trained initial deep learning model.

[0181] An extraction unit is used to extract the encoder in the trained initial deep learning model to obtain the feature extraction model in the target deep learning model.

[0182] Optional, Figure 6 The device may also include:

[0183] The feature data sample generation module is used to process the first graph data sample and the second graph data sample using the feature extraction model in the target deep learning model to obtain the feature data sample of the business entity sample.

[0184] The second label determination module is used to determine label data carried by one or more characteristic data samples of the business entity samples; the label data is used to reflect the benchmark business processing results of a preset category related to the one or more business entity samples.

[0185] The training module may further include:

[0186] The third training unit is used to train the initial classification model using the feature data samples of the one or more business entity samples carrying the label data to obtain predicted business processing results of preset categories related to the one or more business entity samples.

[0187] The third optimization unit is used to adjust the parameters of the initial classification model by taking minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal to obtain the target classification model.

[0188] The model generation unit is used to generate a target deep learning model according to the feature extraction model and the target classification model.

[0189] Optionally, the target deep learning model may also include: a target decoder; the target decoder is connected to the target classification model.

[0190] The third training unit can be specifically used to: use the characteristic data samples of the one or more business entity samples carrying the label data to train the model obtained by connecting the initial decoder and the initial classification model, so as to obtain the predicted business processing results of the preset category related to the one or more business entity samples.

[0191] The third optimization unit can be specifically used to: take minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal, adjust the parameters of the model obtained by connecting the initial decoder and the initial classification model, and obtain the target decoder and the target classification model.

[0192] The model generation unit can be specifically used to generate the target deep learning model based on the feature extraction model, the target decoder and the target classification model.

[0193] Optionally, the graph neural network model can be a model built based on an efficient Transformer architecture for processing graph data; the input of the graph neural network model is the initial feature data of the business entity sample in the first graph data sample and the adjacency relationship data between the business entity samples, and the output of the graph neural network model is the first feature data of the business entity sample.

[0194] Optionally, the hypergraph neural network model may include at least one hypergraph neural network module built on an efficient Transformer architecture for processing hypergraph data, and a feature fusion module; and the hypergraph neural network module is connected to the feature fusion module; for any one of the hypergraph neural network modules, the input of the hypergraph neural network module is the initial feature data of each of the business entity samples contained in a hyperedge in the second graph data sample, and the output of the hypergraph neural network module is the third feature data of the business entity sample contained in the hyperedge; the feature fusion sub-model performs feature fusion processing on the third feature data of the same business entity sample received, and outputs the second feature data of the business entity sample.

[0195] Among them, the computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture.

[0196] Optionally, the graph neural network model may include multiple first network layers connected in sequence and built based on the efficient Transformer architecture; wherein, after the first first network layer receives the initial feature data and the adjacency relationship data of the business entity sample, it performs feature update processing on the business entity sample and outputs the first updated feature data of the business entity sample, and the output of the previous first network layer and the adjacency relationship data are used as inputs for the next first network layer, and the output of the last first network layer is the first feature data of the business entity sample.

[0197] Optionally, the hypergraph neural network module may include multiple second network layers connected in sequence and built based on the efficient Transformer architecture; wherein, after the first second network layer receives the initial feature data of each of the business entity samples contained in a hyperedge, it performs feature update processing on the business entity samples, outputs the second updated feature data of the business entity samples, and the output of the previous second network layer serves as the input of the next second network layer, and the output of the last second network layer is the third feature data of the business entity samples.

[0198] Optionally, the efficient Transformer architecture may include: any one of FastTransformer, LinearTransformer, Performer, Linformer, and Reformer.

[0199] Optionally, the first network layer can be specifically used to: for any one of the business entity samples, use the Gumbel-Softmax algorithm to filter out a first neighbor node from the neighbor nodes of the business entity sample, so as to perform feature update processing on the business entity sample according to feature data of the first neighbor node.

[0200] Optionally, the second network layer can be specifically used to: for any one of the business entity samples contained in a hyperedge, filter out a second neighbor node from the neighbor nodes of the business entity sample using the Gumbel-Softmax algorithm, so as to perform feature update processing on the business entity sample based on the feature data of the second neighbor node.

[0201] The above device embodiments correspond to the method embodiments. For specific descriptions, please refer to the description of the method embodiments, which will not be repeated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For specific descriptions, please refer to the corresponding method embodiments.

[0202] The present specification also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the above Figure 2 The graph data processing method shown and Figure 4 The specific execution process of the model training method shown can be found in the specific description of the relevant embodiments of the above method, which will not be repeated here.

[0203] The embodiments of this specification also provide Figure 7 The structural diagram of the electronic device shown in FIG. Figure 7 At the hardware level, the electronic device may include a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 2 The graph data processing method shown and Figure 4 The specific execution process of the model training method shown can be found in the specific description of the relevant embodiments of the above method, which will not be repeated here.

[0204] Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the executor of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0205] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for computer-readable storage media, computer program products and Figure 6 As for the electronic device shown, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0206] In the 1990s, improvements to a technology could be clearly distinguished as hardware improvements (for example, improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the method flow). However, with the development of technology, many improvements to the method flow today can be regarded as direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that an improvement in a method flow cannot be implemented using a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming it themselves, without having to ask a chip manufacturer to design and produce a dedicated integrated circuit chip. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0207] The controller can be implemented in any appropriate manner, for example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) that can be executed by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art also know that in addition to implementing the controller in a purely computer-readable program code manner, the controller can be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, this controller can be considered as a hardware component, and the devices included therein for implementing various functions can also be regarded as structures within the hardware component. Or even, the devices for implementing various functions can be regarded as both software modules for implementing the method and structures within the hardware component.

[0208] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0209] For the convenience of description, the above device is described in various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0210] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0211] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0212] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0213] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0214] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0215] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0216] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0217] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0218] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0219] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0220] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0221] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.

Claims

1. A graph data processing method, the method comprising: Obtaining first graph data of an association relationship graph of a business entity; The nodes in the association relationship graph are the business entities, and the edges in the association relationship graph are used to reflect the first association relationship between the two nodes connected by the edge; Acquire second graph data of an association relationship hypergraph of the business entity; the nodes in the association relationship hypergraph are the business entities, and the hyperedges in the association relationship hypergraph are used to reflect the second association relationship between the nodes in the hyperedges; Using a target deep learning model, processing the first graph data and the second graph data to obtain target feature data of the business entity; The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model; The graph neural network model is used to generate first feature data of the business entity based on the first graph data, the hypergraph neural network model is used to generate second feature data of the business entity based on the second graph data, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain the target feature data of the business entity.

2. According to the method of claim 1, the graph neural network model is a model for processing graph data based on an efficient Transformer architecture; The input of the graph neural network model is the initial feature data of the business entity in the first graph data and the adjacency relationship data between the business entities, and the output of the graph neural network model is the first feature data of the business entity; in, The computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture.

3. According to the method of claim 2, the graph neural network model comprises a plurality of first network layers connected in sequence and built based on the efficient Transformer architecture; in, After the first first network layer receives the initial feature data and the adjacency relationship data of the business entity, it performs feature update processing on the business entity and outputs the first updated feature data of the business entity. The output of the previous first network layer and the adjacency relationship data serve as input of the next first network layer, and the output of the last first network layer is the first feature data of the business entity.

4. The method according to claim 3, wherein the first network layer is specifically used for: For any of the business entities, a first neighboring node is screened out from the neighboring nodes of the business entity using a Gumbel-Softmax algorithm, so as to perform feature update processing on the business entity according to feature data of the first neighboring node.

5. According to the method of claim 1, the hypergraph neural network model comprises at least one hypergraph neural network module based on an efficient Transformer architecture for processing hypergraph data, and a feature fusion module; in, The hypergraph neural network module is connected to the feature fusion module; For any one of the hypergraph neural network modules, the input of the hypergraph neural network module is the initial feature data of each of the business entities contained in one of the hyperedges in the second graph data, and the output of the hypergraph neural network module is the third feature data of the business entity contained in the hyperedge; The feature fusion sub-model performs feature fusion processing on the received third feature data of the same business entity, and outputs the second feature data of the business entity; The computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture.

6. According to the method of claim 5, the hypergraph neural network module comprises a plurality of second network layers connected in sequence and built based on the efficient Transformer architecture; in, After the first second network layer receives the initial feature data of each business entity contained in a hyperedge, it performs feature update processing on the business entity and outputs the second updated feature data of the business entity. In addition, the output of the previous second network layer serves as the input of the next second network layer, and the output of the last second network layer is the third feature data of the business entity.

7. The method according to claim 6, wherein the second network layer is specifically used for: For any one of the business entities included in a hyperedge, a second neighboring node is screened out from the neighboring nodes of the business entity using the Gumbel-Softmax algorithm, so as to perform feature update processing on the business entity according to feature data of the second neighboring node.

8. The method according to any one of claims 2 to 7, wherein the efficient Transformer architecture comprises: Any of LinearTransformer, FastTransfomer, Performer, Linformer, Reformer.

9. The method according to claim 8, wherein the target deep learning model comprises: Multiple feature extraction models connected in sequence; Wherein, for any one of the feature extraction models, the feature fusion model contained in the feature extraction model is respectively connected with one of the graph neural network models and one of the hypergraph neural network models; After the first feature extraction model receives the first graph data and the second graph data, it performs feature update processing on the business entity and outputs the third updated feature data of the business entity. In addition, the output of the previous feature extraction model serves as the input of the next feature extraction model, and the output of the last feature extraction model is the target feature data of the business entity.

10. The method according to claim 1, wherein the target deep learning model further comprises: Target classification model; The feature fusion model is connected to the target classification model; The processing of the first graph data and the second graph data by using the target deep learning model further includes: The target classification model is used to process the target feature data of one or more of the business entities output by the feature fusion model to obtain business processing results of preset categories related to the one or more business entities.

11. The method according to claim 10, wherein the target deep learning model further comprises: Target decoder; The target decoder is connected to the feature fusion model and the target classification model respectively; After receiving the target feature data of the business entity output by the feature fusion model, the target decoder outputs feature data of preset dimensions of one or more business entities to the target classification model; The target classification model is specifically used to process feature data of preset dimensions of the one or more business entities to obtain business processing results of preset categories related to the one or more business entities.

12. The method according to claim 10 or 11, wherein the business processing results of the preset category include: At least one of a classification result of whether the business entity belongs to a preset category, and an identification result of whether a plurality of the business entities have an association relationship of a preset category.

13. A model training method, the method comprising: Obtain a first graph data sample of an association relationship graph sample of a business entity sample; The nodes in the association relationship graph sample are the business entity samples, and the edges in the association relationship graph sample are used to reflect the first association relationship between the two nodes connected by the edge; Obtain a second graph data sample of an association relationship hypergraph sample of the business entity sample; the nodes in the association relationship hypergraph sample are the business entity samples, and the hyperedges in the association relationship hypergraph sample are used to reflect the second association relationship between the nodes in the hyperedges; Using the first graph data sample and the second graph data sample to train an initial deep learning model to obtain a target deep learning model; The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model; The graph neural network model is used to generate first feature data of the business entity sample based on the first graph data sample, the hypergraph neural network model is used to generate second feature data of the business entity sample based on the second graph data sample, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity sample to obtain target feature data of the business entity sample.

14. The method according to claim 13, wherein the target deep learning model further comprises: Target classification model; The feature fusion model is connected to the target classification model; the method further comprises: Determining label data that the first graph data sample and the second graph data sample need to carry; the label data is used to reflect the benchmark business processing results of a preset category related to one or more of the business entity samples; The using the first graph data sample and the second graph data sample to train the initial deep learning model to obtain a target deep learning model includes: The initial deep learning model is trained using the first graph data sample and the second graph data sample carrying the label data to obtain a predicted business processing result of a preset category related to the one or more business entity samples; wherein the predicted business processing result is obtained by the target classification model performing classification processing on the target feature data of the one or more business entity samples output by the feature fusion model; Taking minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal, the parameters of the initial deep learning model are adjusted to obtain the target deep learning model.

15. The method according to claim 14, wherein the target deep learning model further comprises: Target decoder; The target decoder is connected to the feature fusion model and the target classification model respectively; After receiving the target feature data of the business entity samples output by the feature fusion model, the target decoder outputs the feature data of the preset dimension of the one or more business entity samples to the target classification model; The target classification model is specifically used to perform classification processing on the feature data of preset dimensions of the one or more business entity samples to obtain the predicted business processing result.

16. The method according to claim 13, wherein the target deep learning model comprises: Feature extraction model; The feature fusion model included in the feature extraction model is respectively connected to one of the graph neural network models and one of the hypergraph neural network models; The initial deep learning model is an autoencoder including an encoder and a decoder, and the encoder is the feature extraction model to be trained; The using the first graph data sample and the second graph data sample to train the initial deep learning model to obtain a target deep learning model includes: Using the first graph data sample and the second graph data sample to train the initial deep learning model, and obtain reconstructed feature data of the business entity sample output by the initial deep learning model; Taking minimizing the difference between the initial feature data and the reconstructed feature data of the business entity samples in the first graph data sample and the second graph data sample as a training goal, adjusting the parameters of the initial deep learning model to obtain the trained initial deep learning model; Extract the encoder in the trained initial deep learning model to obtain the feature extraction model in the target deep learning model.

17. The method according to claim 16, wherein the target deep learning model further comprises: Target classification model; The method further comprises: Using the feature extraction model in the target deep learning model, the first graph data sample and the second graph data sample are processed to obtain feature data samples of the business entity sample; Determine label data carried by one or more characteristic data samples of the business entity samples; the label data is used to reflect the benchmark business processing results of a preset category related to the one or more business entity samples; Using the characteristic data samples of the one or more business entity samples carrying the label data, an initial classification model is trained to obtain a predicted business processing result of a preset category related to the one or more business entity samples; Taking minimizing the difference between the predicted business processing result and the benchmark business processing result as a training goal, adjusting the parameters of the initial classification model to obtain the target classification model; The target deep learning model is generated according to the feature extraction model and the target classification model.

18. The method according to claim 17, wherein the target deep learning model further comprises: Target decoder; The target decoder is connected to the target classification model; The using the characteristic data samples of the one or more business entity samples carrying the label data to train the initial classification model to obtain the predicted business processing results of the preset category related to the one or more business entity samples specifically includes: Using the characteristic data samples of the one or more business entity samples carrying the label data, a model obtained by connecting the initial decoder and the initial classification model is trained to obtain a predicted business processing result of a preset category related to the one or more business entity samples; The step of adjusting the parameters of the initial classification model with minimizing the difference between the predicted business processing result and the benchmark business processing result as the training goal to obtain the target classification model specifically includes: Taking minimizing the difference between the predicted service processing result and the benchmark service processing result as a training goal, adjusting parameters of the model obtained by connecting the initial decoder and the initial classification model to obtain the target decoder and the target classification model; The generating the target deep learning model according to the feature extraction model and the target classification model specifically includes: The target deep learning model is generated according to the feature extraction model, the target decoder and the target classification model.

19. According to the method according to any one of claims 13 to 18, the graph neural network model is a model for processing graph data based on an efficient Transformer architecture; The input of the graph neural network model is the initial feature data of the business entity sample in the first graph data sample and the adjacency relationship data between the business entity sample, and the output of the graph neural network model is the first feature data of the business entity sample; The hypergraph neural network model includes at least one hypergraph neural network module built on an efficient Transformer architecture for processing hypergraph data, and a feature fusion module; Furthermore, the hypergraph neural network module is connected to the feature fusion module; For any one of the hypergraph neural network modules, the input of the hypergraph neural network module is the initial feature data of each of the business entity samples contained in one of the hyperedges in the second graph data sample, and the output of the hypergraph neural network module is the third feature data of the business entity sample contained in the hyperedge; The feature fusion sub-model performs feature fusion processing on the received third feature data of the same business entity sample, and outputs the second feature data of the business entity sample; Among them, the computational complexity of the efficient Transformer architecture is less than the computational complexity of the Transformer architecture.

20. According to the method of claim 19, the graph neural network model comprises a plurality of first network layers connected in sequence and built based on the efficient Transformer architecture; in, After receiving the initial feature data and the adjacency relationship data of the business entity sample, the first first network layer performs feature update processing on the business entity sample and outputs the first updated feature data of the business entity sample, and the output of the previous first network layer and the adjacency relationship data serve as the input of the next first network layer, and the output of the last first network layer is the first feature data of the business entity sample; The hypergraph neural network module includes a plurality of second network layers connected in sequence and built based on the efficient Transformer architecture; Wherein, after receiving the initial feature data of each of the business entity samples contained in a hyperedge, the first second network layer performs feature update processing on the business entity sample, outputs the second updated feature data of the business entity sample, and the output of the previous second network layer is used as the input of the next second network layer, and the output of the last second network layer is the third feature data of the business entity sample; Among them, the efficient Transformer architecture includes: any one of FastTransformer, Linear Transformer, Performer, Linformer, and Reformer.

21. The method according to claim 20, wherein the first network layer is specifically used to: for any one of the business entity samples, use the Gumbel-Softmax algorithm to filter out a first neighboring node from the neighboring nodes of the business entity sample, so as to perform feature update processing on the business entity sample according to feature data of the first neighboring node; The second network layer is specifically used for: for any one of the business entity samples contained in a hyperedge, using the Gumbel-Softmax algorithm to filter out a second neighboring node from the neighboring nodes of the business entity sample, so as to perform feature update processing on the business entity sample according to the feature data of the second neighboring node.

22. A graph data processing device, comprising: A first acquisition module, used to acquire first graph data of an association relationship graph of a business entity; The nodes in the association relationship graph are the business entities, and the edges in the association relationship graph are used to reflect the first association relationship between the two nodes connected by the edge; A second acquisition module, used to acquire second graph data of the association relationship hypergraph of the business entity; The nodes in the association relationship hypergraph are the business entities, and the hyperedges in the association relationship hypergraph are used to reflect the second association relationship between the nodes in the hyperedges; A graph data processing module, configured to process the first graph data and the second graph data using a target deep learning model to obtain target feature data of the business entity; The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model; The graph neural network model is used to generate first feature data of the business entity based on the first graph data, the hypergraph neural network model is used to generate second feature data of the business entity based on the second graph data, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity to obtain the target feature data of the business entity.

23. A model training device, comprising: A first acquisition module, used to acquire a first graph data sample of an association relationship graph sample of a business entity sample; The nodes in the association relationship graph sample are the business entity samples, and the edges in the association relationship graph sample are used to reflect the first association relationship between the two nodes connected by the edge; A second acquisition module is used to acquire a second graph data sample of an association relationship hypergraph sample of the business entity sample; the nodes in the association relationship hypergraph sample are the business entity samples, and the hyperedges in the association relationship hypergraph sample are used to reflect the second association relationship between the nodes in the hyperedges; A training module, used to train an initial deep learning model using the first graph data sample and the second graph data sample to obtain a target deep learning model; The target deep learning model includes: a graph neural network model, a hypergraph neural network model and a feature fusion model; the graph neural network model and the hypergraph neural network model are respectively connected to the feature fusion model; The graph neural network model is used to generate first feature data of the business entity sample based on the first graph data sample, the hypergraph neural network model is used to generate second feature data of the business entity sample based on the second graph data sample, and the feature fusion model is used to perform feature fusion processing on the first feature data and the second feature data of the same business entity sample to obtain target feature data of the business entity sample.

24. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 21.

25. An electronic device comprising: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the method as claimed in any one of claims 1 to 21.

Images