Graph neural network training method, apparatus and device, and readable storage medium

Abstract

The invention relates to a graph neural network training method and device, equipment and a readable storage medium. The method comprises the following steps: acquiring a plurality of training edge sets of a graph structure and a to-be-trained graph neural network; determining a neighbor node set of each node in each edge of each training edge set; the method comprises the following steps: determining neighbor position coding information and structure position coding information of each neighbor node in a neighbor node set, and expanding common neighbors among the nodes to obtain common neighbor coding information; according to the structure position coding information and the common neighbor coding information, training the graph neural network, and when a preset training condition is satisfied, ending the training to obtain a trained graph neural network; and obtaining a to-be-processed graph structure, and inputting the to-be-processed graph structure into the trained graph neural network to obtain a prediction result of the to-be-processed graph structure. By adopting the method, the prediction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a graph neural network training method, apparatus, device and readable storage medium. Background Art

[0002] Graph Neural Network (GNN) refers to an algorithm that uses neural networks to learn graph structured data, extract and discover features and patterns in graph structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. Graph neural networks can be widely used in different fields, such as circuit design, traffic flow prediction, traffic route planning, and social impact prediction.

[0003] In related technologies, the training method of graph neural networks limits the ability to obtain more graph features, resulting in low prediction accuracy. Summary of the invention

[0004] Based on this, it is necessary to provide a graph neural network training method, device, computer equipment, computer-readable storage medium and computer program product that can improve prediction accuracy in response to the above-mentioned technical problems.

[0005] In a first aspect, the present application provides a graph neural network training method, comprising:

[0006] Obtain multiple training edge sets of a graph structure and a graph neural network to be trained. The graph structure is determined according to the business scenario. The nodes of the graph structure correspond to the business nodes of the business scenario. The nodes are connected by edges, and the edges are used to represent the business information between the business nodes.

[0007] For each training edge set, determine the neighbor node set of each node in each edge;

[0008] Determine the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure;

[0009] Determine the number of occurrences of each neighbor node in the N neighbor node sets, expand the common neighbors between nodes according to the number of occurrences and neighbor position encoding information, and obtain common neighbor encoding information;

[0010] The graph neural network is trained according to the structural position encoding information and the common neighbor encoding information. When the preset training conditions are met, the training is terminated to obtain a trained graph neural network.

[0011] Obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

[0012] In one embodiment, the business scenario includes at least one of a traffic prediction scenario, a course selection scenario, and a social network analysis scenario.

[0013] When the business scenario is a traffic prediction scenario, the graph structure is determined based on the scheduled operation route information of the transportation tool, the business node of the business scenario includes the target object associated with the scheduled operation route information, and the business information between the business nodes includes the scheduled operation route between the target objects;

[0014] In the scenario where the business scenario is a course selection scenario, the graph structure is determined based on the access behavior data of the selection object and the selected object. The business nodes of the business scenario include the selection object and the selected object, and the business information between the business nodes includes the access behavior between the selection object and the selected object.

[0015] When the business scenario is a social network analysis scenario, the graph structure is determined by the interactive behavior data in the social network, the business nodes of the business scenario include the target objects that generate the interactive behavior data, and the business information between the business nodes includes the interactive behaviors between the target objects.

[0016] In one embodiment, determining the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure includes:

[0017] According to the time sequence information of each neighbor node in the neighbor node set, each neighbor node is sorted in chronological order, and the sorted neighbor nodes are encoded to obtain the neighbor position encoding information of each neighbor node;

[0018] Determine the number of paths between each neighbor node and the node;

[0019] The structural position encoding information of the neighbor nodes in the graph structure is determined according to the number of paths.

[0020] In one embodiment, for each training edge set, determining a neighbor node set of each node in each edge includes:

[0021] For each training edge set, the neighbor nodes of each node in each edge are sampled in parallel to obtain the first-order neighbor nodes and high-order neighbor nodes of each node;

[0022] The neighbor node set of each node is obtained according to the first-order neighbor nodes and high-order neighbor nodes.

[0023] In one embodiment, the nodes include a source node and a target node, the number of occurrences of each neighbor node in the N neighbor node sets is determined, and the common neighbors between the nodes are expanded according to the number of occurrences and the neighbor position encoding information to obtain the common neighbor encoding information, including:

[0024] Determine a common neighbor set of a first neighbor node set of a source node and a second neighbor node set of a target node;

[0025] For each neighbor node in the first neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding first two-dimensional vector;

[0026] For each neighbor node in the second neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding second two-dimensional vector;

[0027] Performing Fourier transform on the first two-dimensional vector and the second two-dimensional vector respectively to obtain first embedded representation and second embedded representation corresponding to each other;

[0028] According to the first embedding representation, the second embedding representation and the neighbor position encoding information, information of each common neighbor in the common neighbor set is determined to obtain common neighbor encoding information.

[0029] In one embodiment, training a graph neural network according to structural position encoding information and common neighbor encoding information includes:

[0030] Get the original features of each node, the edge feature encoding information and time encoding information of each edge;

[0031] The structural position encoding information is concatenated with the corresponding original features to obtain the node encoding information of each node;

[0032] The node coding information, edge feature coding information, time coding information and common neighbor coding information are concatenated and mapped to obtain mapping features;

[0033] The Transformer of the graph neural network is trained using the mapping features.

[0034] In one embodiment, obtaining multiple training edge sets of a graph structure includes:

[0035] If the current training of the graph neural network is the first training, all edge sets corresponding to the graph structure are divided based on the normal distribution to obtain multiple training edge sets;

[0036] For each training from the start of the second training to the end of the i-th training, obtain the previous training result of the graph neural network;

[0037] Update the weight of each edge in the graph structure according to the last training result to obtain the updated weight of each edge;

[0038] According to the update weight of each edge, the target edge is extracted from all edge sets corresponding to the graph structure to obtain multiple training edge sets.

[0039] In one embodiment, the method further comprises:

[0040] Processing the pre-stored graph structure to obtain a graph data structure storing connection relationships between nodes in the graph structure;

[0041] Split the graph data structure to obtain all edge sets of the graph structure.

[0042] In a second aspect, the present application also provides a graph neural network training device, comprising:

[0043] The data acquisition module is used to obtain multiple training edge sets of the graph structure and the graph neural network to be trained. The graph structure is determined according to the business scenario. The nodes of the graph structure correspond to the business nodes of the business scenario. The nodes are connected by edges, and the edges are used to represent the business information between the business nodes.

[0044] A neighbor node processing module, used for determining, for each training edge set, a neighbor node set of each node in each edge;

[0045] A first encoding module is used to determine the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure;

[0046] The second encoding module is used to determine the number of occurrences of each neighbor node in the N neighbor node sets, and expand the common neighbors between the nodes according to the number of occurrences and the neighbor position encoding information to obtain the common neighbor encoding information;

[0047] A training module is used to train the graph neural network according to the structural position encoding information and the common neighbor encoding information, and when the preset training conditions are met, the training is terminated to obtain a trained graph neural network;

[0048] The prediction module is used to obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the target operation route information in the graph structure to be processed.

[0049] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the following steps when executing the computer program: obtaining multiple training edge sets of a graph structure and a graph neural network to be trained, obtaining multiple training edge sets of a graph structure and a graph neural network to be trained, wherein the graph structure is determined according to a business scenario, the nodes of the graph structure correspond to the business nodes of the business scenario, the nodes are connected by edges, and the edges are used to represent the business information between the business nodes;

[0050] For each training edge set, determine the neighbor node set of each node in each edge;

[0051] Determine the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure;

[0052] Determine the number of occurrences of each neighbor node in the N neighbor node sets, expand the common neighbors between nodes according to the number of occurrences and neighbor position encoding information, and obtain common neighbor encoding information;

[0053] The graph neural network is trained according to the structural position encoding information and the common neighbor encoding information. When the preset training conditions are met, the training is terminated to obtain a trained graph neural network.

[0054] Obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

[0055] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the following steps are implemented:

[0056] Obtain multiple training edge sets of a graph structure and a graph neural network to be trained. Obtain multiple training edge sets of a graph structure and a graph neural network to be trained. The graph structure is determined according to the business scenario. The nodes of the graph structure correspond to the business nodes of the business scenario. The nodes are connected by edges, and the edges are used to represent the business information between the business nodes.

[0057] For each training edge set, determine the neighbor node set of each node in each edge;

[0058] Determine the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure;

[0059] Determine the number of occurrences of each neighbor node in the N neighbor node sets, expand the common neighbors between nodes according to the number of occurrences and neighbor position encoding information, and obtain common neighbor encoding information;

[0060] The graph neural network is trained according to the structural position encoding information and the common neighbor encoding information. When the preset training conditions are met, the training is terminated to obtain a trained graph neural network.

[0061] Obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

[0062] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0063] Obtain multiple training edge sets of a graph structure and a graph neural network to be trained. Obtain multiple training edge sets of a graph structure and a graph neural network to be trained. The graph structure is determined according to the business scenario. The nodes of the graph structure correspond to the business nodes of the business scenario. The nodes are connected by edges, and the edges are used to represent the business information between the business nodes.

[0064] For each training edge set, determine the neighbor node set of each node in each edge;

[0065] Determine the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure;

[0066] Determine the number of occurrences of each neighbor node in the N neighbor node sets, expand the common neighbors between nodes according to the number of occurrences and neighbor position encoding information, and obtain common neighbor encoding information;

[0067] The graph neural network is trained according to the structural position encoding information and the common neighbor encoding information. When the preset training conditions are met, the training is terminated to obtain a trained graph neural network.

[0068] Obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

[0069] The above-mentioned graph neural network training method, apparatus, computer equipment, computer-readable storage medium and computer program product obtain a graph structure determined based on a business scenario, and split the graph structure into multiple training edge sets for training the graph neural network according to the characteristics of the graph structure and the dependencies between nodes, that is, use small batches of training samples to train the graph neural network to be trained. On this basis, by considering the position of each edge node in the graph structure and the position of the node in the neighboring nodes in the training edge set, the corresponding neighbor position coding information and structure position coding information are obtained. The neighbor position coding information and structure position coding information are added on the basis of considering time coding, and the enhanced input features are In order to further improve the structure representation ability of the graph neural network, the common neighbors between nodes are expanded to capture more frequency domain information and obtain the common neighbor encoding information. In this way, the information of the input features is enhanced to train the time-series graph neural network, which solves the missing structural information in the model and improves the representation ability of the time-series graph neural network for structural sensitivity, thereby obtaining a trained graph neural network. Based on the trained graph neural network, when processing the graph structure to be processed, it can better capture the changes of nodes over time and time-sensitive features, and can more comprehensively determine the rationality of the business information in the graph structure to be processed, thereby improving the accuracy of business information prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0071] Figure 1 A diagram showing an application environment of a neural network training method in one embodiment;

[0072] Figure 2 The figure is a schematic diagram of a process flow of a neural network training method in one embodiment;

[0073] Figure 3 is a schematic diagram of a flow chart of a graph neural network to be trained in one embodiment;

[0074] Figure 4 is a schematic flow chart of step 206 in one embodiment;

[0075] Figure 5 A schematic diagram of a process for determining common neighbor coding information in one embodiment;

[0076] Figure 6 is a flow chart of time expenditure in one embodiment;

[0077] Figure 7 is a schematic diagram of a flow chart of a neighbor sampling method in one embodiment;

[0078] Figure 8 A schematic diagram of a detailed flow chart of a feature processing stage in one embodiment;

[0079] Fig. 9 A schematic diagram of a flow chart of a method for determining a training sample in one embodiment;

[0080] Fig.10 FIG. 1 is a flow chart of a neural network training method according to another embodiment;

[0081] Fig.11 It is a structural block diagram of a neural network training device in one embodiment;

[0082] Fig.12 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0084] Graph neural networks can be widely used in different fields, such as circuit design, traffic flow prediction, traffic route planning, and social impact prediction. For example, in a traffic scenario, route conflicts can be understood as mutual interference or potential collisions between vehicles, pedestrians, or other traffic participants during road use. If there is a route conflict, there will be safety hazards and waste of resources, so it is necessary to ensure traffic safety and traffic resource utilization. In related technologies, traditional optimization algorithms are used to ensure the reliability of the route, thereby ensuring traffic safety and traffic resource utilization.

[0085] However, for traffic scenarios with complex relationships and dynamic changes, the optimization algorithms of related technologies cannot accurately predict the rationality of running routes, resulting in low utilization of traffic resources. Therefore, in view of the low prediction accuracy of existing graph neural networks, it is proposed to use training neural networks to improve the prediction accuracy. However, temporal graphs are widely used in predictions in different fields, but in temporal networks, the initial exploration of temporal graph transformers (TGTs) is limited. The related dynamic graph neural networks trained based on the hierarchical aggregation method of graph neural networks bring challenges such as over-smoothing and over-compression. Another neural network architecture, Transformer, has shown excellent performance in modeling long-term dependencies, making it a promising solution for representation learning in temporal networks. Its powerful Transformer model has attracted emerging attempts in temporal networks, which represents node dynamic embedding by encoding training of one-hop neighbors. However, due to the sharp increase in computational complexity caused by neighborhood explosion, these temporal graph transformers (TGTs) usually adopt overly simple neighborhood and location information designs. They only consider temporal encoding without distinguishing the position and neighbor set of nodes in the graph structure. Second, they ignore high-order neighbor information, which limits the model's ability to capture more graph features and leads to low prediction accuracy.

[0086] Therefore, in order to solve the technical problem of low prediction accuracy, a graph neural network training method is proposed. By obtaining the graph structure, multiple training edge sets for training the graph neural network are determined according to the graph structure. For each training edge set, the neighbor node set of each node in each edge is determined. The position of each neighbor node in the graph structure and the position information in the neighbor node set are considered, and the input features are enhanced to train the graph neural network, which enhances the model representation ability and improves the accuracy of training. When processing the graph structure to be processed based on the trained graph neural network, the changes of nodes over time and time-sensitive features can be better captured, and the accuracy of the prediction of the edges to be processed in the graph structure to be processed can be determined more comprehensively. In the traffic prediction scenario, the rationality of the target operation route information can be achieved, and the reasonable planning of the operation route can be realized, ensuring the safety of driving and improving the utilization of traffic resources.

[0087] The graph neural network training method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal obtains multiple training edge sets of the graph structure and the graph neural network to be trained from the server; for each training edge set, determine the neighbor node set of each node in each edge; determine the neighbor position coding information of each neighbor node in the neighbor node set and the structural position coding information of each neighbor node in the graph structure; determine the number of occurrences of each neighbor node in N neighbor node sets, expand the common neighbors between nodes according to the number of occurrences and the neighbor position coding information, and obtain the common neighbor coding information; train the graph neural network according to the structural position coding information and the common neighbor coding information, and when the preset training conditions are met, end the training to obtain the trained graph neural network; obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the target operation route information in the graph structure to be processed.

[0088] Among them, the terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0089] In an exemplary embodiment, Figure 2 As shown, a graph neural network training method is provided, which is applied to Figure 1 The terminal in is taken as an example to illustrate, including the following steps 202 to 212. Among them:

[0090] Step 202, obtain multiple training edge sets of the graph structure and the graph neural network to be trained. The graph structure is determined according to the business scenario. The nodes of the graph structure correspond to the business nodes of the business scenario. The nodes are connected by edges, and the edges are used to represent the business information between the business nodes. Among them, the graph neural network to be trained can be a pre-built Transformer-based fast perception time-series graph neural network, called QSFormer. The schematic diagram of the graph neural network to be trained is as follows Figure 3 As shown, it includes a dynamic graph input module, a feature processing module and a Transformer module. The dynamic graph input module is used to generate small batch samples for training. The small batch samples can be generated using an adaptive method to generate small batches that are beneficial to training. Being beneficial to training can be understood as being able to accelerate the convergence speed of model training and improve training efficiency.

[0091] The feature processing module can be used to accept a mini-batch as input, process the mini-batch on the edge unit, extract high-order neighbors, and encode features for the source node u and the target node v of the edge (u,v,t) at time t, and finally connect all the features to obtain the feature encoding Z_(u,v)^t. That is, the feature processing module at least includes a padded parallel neighbor sampler, an encoder, and a mapper. The encoder may include a position-differentiated encoder, an extended common neighbor encoder, etc. Among them, the encoder can be used to encode the features and related timestamps of nodes u and v and their neighborhoods, edges (u,v,t) into node, edge, time, and extended common neighbor encodings, represented as node encoding information Edge Encoding Information Time code information Extended common neighbor encoding information Where * represents the source node u or the target node v. The time coding function for Δt′=tt′ is given as follows, and the implementation method of the time coding information is obtained:

[0092]

[0093] where ω 1 ,…,ω dt is a trainable parameter, d t is the dimension of time encoding. The mapper can be the encoded features that are then patched, mapped, and concat into a unified feature encoding. Represents the combined state of the edges input to the Transformer module.

[0094] The Transformer module can be an output encoding that receives feature processing. As input, and trained through multiple layers of Multi-Head Attention, Feed Forward, etc., finally, it is passed through an average pooling layer AvgPool to get the final embedding, which is split into the embedding representation of the source node u and the target node v of the edge (u, v, t) and That is, the Transformer module includes normalization layer, multi-head attention layer, feedforward network and average pooling, among which the multi-head attention layer allows the model to focus on different embedded segments and capture the complex dependencies in the graph; the feedforward network is after normalization and residual connection, and the attention output is further refined through the feedforward network; the average pooling layer is the last step of the Transformer, which involves pooling the output to form the representation of the source node u and the target node v of the edge, which can be expressed as This is crucial for downstream tasks such as link prediction. In the link prediction task, we evaluate The similarity between them is used to determine whether there is a link between u and v at time t.

[0095] The graph structure can be understood as a dynamic graph. The graph structure is determined according to the business scenario. The nodes of the graph structure correspond to the business nodes of the business scenario. The nodes are connected by edges, and the edges are used to represent the business information between the business nodes. The business scenario includes at least one of the traffic prediction scenario, the course selection scenario, and the social network analysis scenario. This embodiment takes the traffic prediction scenario as an example. It should be noted that different application scenarios have different corresponding graph structures, and the specific meanings of the nodes in the graph structure and the meanings of the edges between the nodes are also different.

[0096] Optionally, in the case where the business scenario is a traffic prediction scenario, the graph structure is determined based on the scheduled operation route information of the vehicle, the business node of the business scenario includes the target object associated with the scheduled operation route information, and the business information between the business nodes includes the scheduled operation route between the target objects. That is, the graph structure is determined based on the scheduled operation route information of the vehicle, and the scheduled operation route information may be the operation route information between multiple regions within a preset time period. The scheduled operation route information may include the scheduled operation route, the operation time of the scheduled operation route, and the target object associated with the scheduled operation route, as well as the associated information related to the target object, and the associated information includes the airport type, the regional identifier of the area where the airport is located, the airport identifier, and the geographical location of the area where the airport is located.

[0097] The number of target objects includes at least two, and the target object can be an airport across regions. One node corresponds to one target object, and the target object is described as an airport. The node features of the node include the airport type, the regional identifier of the area where the airport is located, the airport identifier, and the geographical location of the area where the airport is located. The edge features corresponding to the edges between nodes include the number of flights and the flight time corresponding to each flight. It should be noted that there may be a transfer airport between the source airport and the destination airport. In this case, the route from the source airport to the destination airport will include at least one path, that is, there will be at least one edge.

[0098] Optionally, in the scenario where the business scenario is a course selection scenario, the graph structure is determined based on the access behavior data of the selection object and the selected object, the business nodes of the business scenario include the selection object and the selected object, and the business information between the business nodes includes the access behavior between the selection object and the selected object, that is, for the large open network course scenario, the graph structure includes two types of nodes, the business nodes corresponding to the nodes can be the selection object (for example, students) and / or the selected object (for example, course units), the edges between the nodes represent the students' access behavior to specific content units, and each link is associated with a multidimensional feature, which can also be called an N-dimensional feature, and N can be, but is not limited to, 4, or 3 or 2. Among them, the association of multidimensional features can be determined by selecting N behavior information from the learning behavior information, and the selection method can be determined according to actual needs, which is not limited here. Among them, learning behavior information may include learning start time (start_time_DI), learning end time (last_event_DI), course interaction times (nevents), learning access days (ndays_act), video play times (nplay_video), learning chapters (nchapters), posting times (nforumposts), etc., among which, course interaction times refer to the number of times learners log in to their learning accounts; posting times refer to the number of times learners speak in discussion areas and forums.

[0099] Optionally, in the case where the business scenario is a social network analysis scenario, the graph structure is determined by the interactive behavior data in the social network, the business nodes of the business scenario include the target objects that generate the interactive behavior data, and the business information between the business nodes includes the interactive behavior between the target objects, that is, for the social network scenario, the interactive behavior data includes the business nodes determined based on the directed reply network information in the social network, the business nodes corresponding to the nodes of the graph structure can be but are not limited to different users and / or commodities, and the business information between the business nodes represented by the edges connecting the nodes may include interactive behaviors between users, for example, a reply from one user to another.

[0100] Optionally, the relevant data for determining the graph structure may be obtained from data pre-stored in a server, and matching target data may be determined from the server according to an identifier carried in a data acquisition instruction.

[0101] Exemplarily, multiple training edge sets of a graph structure determined based on the scheduled operation route information of a transportation tool and a graph neural network to be trained are obtained from a server, wherein each edge in the training edge set has two nodes, which may be referred to as a source node and a target node.

[0102] Furthermore, multiple training edge sets of the graph structure can be obtained by processing a pre-stored graph structure to obtain a graph data structure that stores the connection relationship between nodes in the graph structure; splitting the graph data structure to obtain all edge sets of the graph structure, and generating all edge sets based on a normal distribution.

[0103] Step 204: for each training edge set, determine the neighbor node set of each node in each edge.

[0104] Among them, neighbor nodes include first-order neighbors and high-order neighbors. First-order neighbors can be understood as neighbor nodes directly connected to the node, and high-order neighbors can be understood as neighbor nodes reachable through multi-hop paths. In addition, high-order neighbors can provide richer contextual information, which helps the model capture the complex relationship between nodes. In other words, the embedding features calculated by considering higher-order neighbors contain more information than those calculated by only considering low-order neighbors. In addition, from the perspective of graph structure, the addition of high-order neighbors allows the model to perceive a more comprehensive graph structure. Each node can serve as the root node of all neighbor nodes in its corresponding neighbor node set.

[0105] For example, there are dependencies between the nodes of the graph structure. On the basis of obtaining multiple training edge sets of the graph structure, for each node in each training edge set, according to the dependencies between the nodes, each node in the training edge set is analyzed to obtain the neighbor node set of each node. Among them, the method of determining the neighbor node set of each node can be implemented by existing methods, which will not be described in detail here.

[0106] Step 206: Determine the neighbor position coding information of each neighbor node in the neighbor node set and the structural position coding information of each neighbor node in the graph structure.

[0107] It should be noted that the transformer module does not distinguish the position of the token and requires additional position information. However, in the research of dynamic graph transformers, only time coding is often used to provide the transformer with the timing information of the node, while ignoring the position of the node in the neighbor set and the position in the graph structure, which leads to less information in the input features. In order to enable the model to perceive a more comprehensive graph structure and improve the accuracy of prediction, it is necessary to consider the position of each neighbor node in the neighbor node set and the position in the graph structure.

[0108] Neighbor position encoding information is used to determine the node in the neighbor node set (also called the extended neighborhood set) ), combined with the Transformer's attention mechanism, indirectly provides insights into the underlying graph structure. This encoding scheme generates a unique code for each node, making it possible to distinguish identical nodes that appear in the same neighborhood set. Identical nodes that appear in the same neighborhood set can be understood as taking airport A and airport B as an example. When there are multiple edges between airport A and airport B, the node corresponding to airport B is a neighbor node of the node corresponding to airport A. Then, in the neighbor node set of the node corresponding to airport A, the node corresponding to airport B will appear multiple times. The implementation method of neighbor position encoding information can be to sort each neighbor node in each neighbor node set in chronological order and determine the number of the sorted neighbor nodes.

[0109] The structural position encoding information is used to identify the number of hops between the neighbor node and the root node, which can directly reflect the graph structure information. The number of hops here refers to the number of edges that need to be passed from one node to another in the graph structure, which can also be understood as the number of paths. Since the Transformer model only processes sequence inputs, high-order neighbors are flattened into a sequence. In order to identify the number of hops from the sequence, this method uses One-Hot Code to encode the structural position and obtain the corresponding structural position encoding information.

[0110] Step 208, determining the number of occurrences of each neighbor node in the N neighbor node sets, and expanding the common neighbors between the nodes according to the number of occurrences and the neighbor position coding information to obtain the common neighbor coding information.

[0111] It should be noted that the common neighbor encoding information can also be called the extended common neighbor encoding information. The original common neighbors are the intersection of the first-order neighbor sets of the two nodes. On the basis of introducing high-order neighbors, the common neighbors need to be expanded to the intersection of the high-order neighbor sets of the two nodes, that is, the extended common neighbor set includes the intersection of the first-order neighbor sets of the two nodes and the intersection of the high-order neighbor sets of the two nodes. N is a positive integer, and N can be but is not limited to 2.

[0112] Optionally, the information of the extended common neighbors that enumerates the frequency of occurrence of each neighbor in the N neighbor node sets is obtained by encoding. Further, taking N as 2, the common neighbor set of the first neighbor node set of the source node and the second neighbor node set of the target node is determined; for each neighbor node in the first neighbor node set of the source node, the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node is determined to obtain the corresponding first two-dimensional vector; for each neighbor node in the second neighbor node set of the source node, the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node is determined to obtain the corresponding second two-dimensional vector; Fourier transform is performed on the first two-dimensional vector and the second two-dimensional vector respectively to obtain the first embedded representation and the second embedded representation corresponding to each; according to the first embedded representation, the second embedded representation and the neighbor position encoding information, the information of each common neighbor in the common neighbor set is determined to obtain the common neighbor encoding information.

[0113] Step 210, training the graph neural network according to the structural position encoding information and the common neighbor encoding information, and when the preset training conditions are met, ending the training to obtain a trained graph neural network.

[0114] Among them, the preset training conditions can be understood as model convergence conditions.

[0115] Optionally, for each edge in the training edge set, the source node u and the target node v of the edge (u, v, t) at time t are encoded features to obtain the corresponding time coding information, edge coding information and node coding information of the source node / target node, and all features such as the time coding information, edge coding information, node coding information of the source node / target node, structural position coding information and common neighbor coding information are connected to obtain splicing features, and the obtained splicing features are used to train the graph neural network. When the preset training conditions are met, the training is terminated to obtain a trained graph neural network.

[0116] Step 212, obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

[0117] The structure to be processed includes a newly added edge structure, and the edge structure may be a route newly added between Airport A and Airport B at time t.

[0118] Optionally, the graph structure to be processed is input into a trained graph neural network, and the predicted structure of the edges to be processed in the graph structure to be processed can be predicted based on the trained graph neural network, that is, it can be judged whether the newly added operation route is reasonable. Compared with the existing method, the correlation information between the graph structures formed by the flight operation routes can be fully considered, which improves the accuracy of the prediction of the newly added operation routes, thereby avoiding the waste of traffic resources and ensuring the safety of flights. Furthermore, the graph neural network trained in the above manner can accurately determine the relationship between users when predicting the graph structure to be processed in the social network analysis scenario. In addition, data can be pushed to users, such as friend push, message push, etc., thereby improving the accuracy of data push. In the scenario of course selection, when predicting based on the graph structure to be processed, the matching degree between students and course units, as well as the attention of course units, etc. can be accurately determined, thereby optimizing the allocation of resources.

[0119] In the above-mentioned graph neural network training method, by obtaining a graph structure representing the scheduled operating route information of a transportation tool, the graph structure is split into multiple training edge sets for training the graph neural network according to the characteristics of the graph structure and the dependencies between nodes. That is, the graph neural network to be trained is trained using small batches of training samples. On this basis, by considering the position of each edge node in the graph structure and the position of the node in the neighboring nodes in the training edge set, the corresponding neighbor position coding information and structure position coding information are obtained. The neighbor position coding information and structure position coding information are added on the basis of considering time coding, thereby enhancing the amount of information of the input features. In order to further improve the representation of the structure by the graph neural network, The representation ability is improved by expanding the common neighbors between nodes, capturing more frequency domain information, and obtaining the common neighbor encoding information. In this way, the information of the input features is enhanced to train the time-series graph neural network, which solves the missing structural information in the model and improves the representation ability of the time-series graph neural network for structural sensitivity, thereby obtaining a trained graph neural network. When processing the graph structure to be processed based on the trained graph neural network, it can better capture the changes of nodes over time and time-sensitive features, and can more comprehensively determine the rationality of the target operation route information in the graph structure to be processed, realize the reasonable planning of the operation route, ensure driving safety, and improve the utilization rate of traffic resources.

[0120] Existing TGT models mainly focus on temporal encoding, but they often ignore the key role of graph structure. This oversight may limit the model's ability to capture structural differences between adjacent nodes, thereby affecting the accuracy of training. Therefore, in an exemplary embodiment, Figure 4 As shown, step 206 includes steps 402 to 406. Among them:

[0121] Step 402, sorting each neighbor node in chronological order according to the timing information of each neighbor node in the neighbor node set, encoding the sorted neighbor nodes, and obtaining neighbor position encoding information of each neighbor node.

[0122] The timing information can be understood as the interaction time between the neighbor node and the root node. This interaction time can be the flight time of the flight, and the flight time can be the take-off time and / or landing time. Consider a sorted neighbor list in and This list represents the neighbors of a node v at time t. Each neighbor u i With a time t i Associated, indicating that v and u i The interaction time between neighbors. The neighbor position identity encoding of each neighbor is defined as follows:

[0123]

[0124] Step 404, determining the number of paths from each neighbor node to the root node.

[0125] Among them, the number of paths between each neighbor node and the root node can be determined according to the dependency relationship between the nodes, and the number of paths can also be called the number of hops.

[0126] Step 406, determining the structural position coding information of the neighbor node in the graph structure according to the number of paths.

[0127] Where k is used to represent the number of hops, and the structure position identification code is defined as follows:

[0128]

[0129] Among them, p represents the path between the source node u and the target node v, e is the edge, and u i Represents a neighbor node.

[0130] In the above embodiment, by combining the location information and the structure information to encode the neighbor nodes in the graph, the model can perceive a more comprehensive graph structure, thereby improving the model's prediction performance for graph data. In addition, by effectively encoding the information of the neighbor nodes, the amount of data that the model needs to process during training and reasoning can be reduced, thereby improving the algorithm's computational efficiency and reducing the consumption of computing resources.

[0131] It should be noted that after the high-order neighbors are expanded, the common neighbor codes need to be expanded. In an exemplary embodiment, the nodes include a source node and a target node, such as Figure 5As shown, a method for determining common neighbor coding information is provided, including steps 502 to 510, wherein:

[0132] Step 502: determine a common neighbor set of a first neighbor node set of a source node and a second neighbor node set of a target node.

[0133] It should be noted that for an interaction (u, v, t), the original definition of the neighbor node set of node u at time t is For all nodes v that are directly connected to u by edges and whose time is no later than t: In this embodiment, the multi-hop neighbors of the source node u at time t are defined as the extended neighbor set The expanded neighbor set of the target node The determination method of is the same as that of the source node, which will not be repeated here. Then the common neighbor set of the first neighbor node set of the source node and the second neighbor node set of the target node can be expressed as: the extended common neighbor set The endpoint extends the neighbor set.

[0134] Step 504 , for each neighbor node in the first neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding first two-dimensional vector.

[0135] Step 506: for each neighbor node in the second neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding second two-dimensional vector.

[0136] Optionally, the extended neighbor set for the source node For each neighbor in and to obtain the first two-dimensional vector Similarly, for For each neighbor in and The number of occurrences in the corresponding second two-dimensional vector

[0137] Step 508: Perform Fourier transform on the first two-dimensional vector and the second two-dimensional vector respectively to obtain the first embedded representation and the second embedded representation corresponding to each other.

[0138] Among them, since Fourier transform can provide additional attention to frequency domain information to enhance the representation ability of the structure and reduce the time consumption of encoding common neighbors, a non-learning Fourier transform method is defined for the first two-dimensional vector and the second two-dimensional vector to obtain the corresponding first embedding representation and second embedding representation.

[0139] Defining a non-learning Fourier transform method for the first two-dimensional vector and the second two-dimensional vector can be expressed as follows:

[0140]

[0141] Among them, freq(x) represents The frequency count of the xth element of freq Indicates the size of the frequency vector dimension. Where * can be u, v. Apply the CoE function to and To obtain the first embedding representation and the second embedding representation

[0142] Step 510, determining information of each common neighbor in the common neighbor set according to the first embedded representation, the second embedded representation and the neighbor position encoding information, and obtaining common neighbor encoding information.

[0143] Among them, determine the common neighbor coding information, that is, determine the high-order neighbor coding that distinguishes the final position This can be achieved by:

[0144]

[0145] Among them, * can be u, v, t is time, and NE represents the neighbor position encoding information.

[0146] In the above embodiment, a non-learning Fourier transform is used to capture more frequency domain information, thereby further improving the representation capability of the structure. Moreover, this method does not require learning, and it can avoid back-propagation calculations and improve training efficiency.

[0147] It is understandable that if Figure 6As shown, it is a schematic diagram of the time cost process in an embodiment, including loadfeature, encodeCo, neighbor sample, transformer and construct patches, wherein encodeCo represents common neighbor encoding, neighbor sample represents neighbor sampling, and it can be seen that the time cost ring diagram of training on DyGFormer shows that the time cost of neighbor sampling and the time cost of encoding common neighbors occupy a considerable part of its time cost. In order to directly reduce the time cost caused by adding high-order neighbors, the learnable MLP layer (MultilayerPerceptron) for encoding common neighbors is replaced with a non-learning Fourier transform encoding method to reduce the time cost of encoding common neighbors, and in order to speed up neighbor sampling, a filling parallel sampler is proposed to efficiently perform neighbor sampling. In an exemplary embodiment, as shown in Figure 7 As shown, a neighbor sampling method is provided, including steps 702 to 704, wherein:

[0148] Step 702: for each training edge set, sample the neighbor nodes of each node in each edge in parallel to obtain the first-order neighbor nodes and high-order neighbor nodes of each node.

[0149] Among them, the search method can be to first initialize an array filled with zeros neigh for sampling results, fill unfilled positions with zeros, and optimize the array for parallel operation and memory block copying.

[0150] Step 704: Obtain a neighbor node set of each node according to the first-order neighbor nodes and the high-order neighbor nodes.

[0151] Exemplarily, the sampling principle of filling the parallel neighbor sampler can be: input a target node list Expansion Factor The graph G in adjacency format,

[0152] Initialize a shape The zero tensor neigh;

[0153] For each expansion factor i from 1 to l: In parallel for each node j from 1 to b:

[0154]

[0155] end←Binary search

[0156] start←max(0,end-m i );

[0157] Memory block copy Output the sampled neighbors neigh. In this way, the neighbor node set of each node can be obtained.

[0158] Optionally, for each training edge set, the corresponding graph is structured in an adjacency format, neighbors are sorted by time, and neighbor information such as node ID and timestamp is accessed in O(1). The corresponding neighbor array of each node, the size of the array can be preset, and binary search is used to quickly find the earliest neighbor before a specific timestamp. Further, memory block copy is used to transfer data from the earliest valid neighbor to the required number of neighbors in neigh. If the existing neighbors are insufficient, this method copies all available neighbors to neigh, which can be understood as the remaining positions of the array neigh are represented by zeros. This method can be specifically implemented by using C++ to implement thread-level parallelism and direct memory operations.

[0159] In this embodiment, a high-order neighbor sampling technique is used to fill parallel acceleration, and a binary search is used to quickly find the earliest neighbor before a specific timestamp to achieve a time complexity of O(logn). In addition, a memory-friendly memory block copy operation is used to maximize CPU parallelism and reduce memory access time.

[0160] The above respectively discusses the specific implementation steps of filling parallelism and encoding. The following is a detailed flow chart of the feature processing stage in an exemplary embodiment. Figure 8 As shown, the high-order neighbors of the node are explained by taking the second order as an example. The feature processing includes three parts: efficiency optimization, perception enhancement and mapping processing. For each training edge set Batch, each edge can be represented as (u, v, t). By filling the parallel sampler, the 2-hop neighbors of u and v, that is, the neighbor node set, are obtained, and the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure are determined; the number of occurrences of each neighbor node in the two neighbor node sets is determined, and the common neighbors between the nodes are expanded according to the number of occurrences and the neighbor position encoding information to obtain the common neighbor encoding information; and the source node u and the target node v of the edge (u, v, t) at time t are encoded features, all the features are connected, and through mapping processing, the feature encoding Z_(u, v)^t is obtained, that is, the encoded features are then patched, mapped and concat into a unified feature encoding. Represents the comprehensive state of the edge input to the Transformer module. This approach improves the model's perception of the graph structure and node positions by increasing the amount of information in the input features.

[0161] The increase in the amount of information brings about an increase in the amount of calculation, which leads to efficiency problems. In order to reduce the efficiency problems caused by the increase in the amount of information, in an exemplary embodiment, Fig. 9As shown, a method for determining a training sample is provided, comprising steps 902 to 908, wherein:

[0162] Step 902: If the current training of the graph neural network is the first training, all edge sets corresponding to the graph structure are divided based on the normal distribution to obtain multiple training edge sets.

[0163] Among them, multiple training edge sets can be understood as being generated based on small batches. It should be noted that during the first training, all the obtained training edge sets are used to train the graph neural network to be trained, and corresponding training results will be obtained.

[0164] Step 904: For each training from the start of the second training to the end of the i-th training, obtain the previous training result of the graph neural network.

[0165] Step 906, update the weight of each edge in the graph structure according to the last training result to obtain the updated weight of each edge.

[0166] The updating method may be based on a preset function. For example, each positive sample e(u i ,v i ,t i )∈E B The importance weight w i =P(e) is updated according to the following function:

[0167]

[0168] in, represents the embedding of the source node u of the edge, represents the embedding of the target node v of the edge, γ is a hyperparameter representing the amplitude of the uniform distribution mixed with the adaptive importance distribution, and Pred is the prediction head of the model, which consists of two layers of MLP and is used to finally give the probability that there is an edge connection between two points.

[0169] Step 908 , extracting target edges from all edge sets corresponding to the graph structure according to the update weight of each edge, to obtain multiple training edge sets.

[0170] For example, a list of importance weights is determined To evaluate the importance level of each edge in the dynamic graph, the list is initialized to a uniform distribution. E_train is the set of training edges. Then, a batch of training edges E are randomly selected according to the corresponding importance score ratio. B After forward propagation, each positive sample e(u i ,v i ,t i )∈E B The importance weight w i=P(e) is updated according to the following function:

[0171] Among them, among them, represents the embedding of the source node u of the edge, Represents the embedding of the target node v of the edge, γ is a hyperparameter representing the amplitude of the uniform distribution mixed with the adaptive importance distribution, and Pred is the prediction head of the model, which consists of two layers of MLP and is used to finally give the probability that an edge connection exists between two points. The first generation of a small batch is based on a normal distribution. In order to accelerate convergence, selecting positive samples with high confidence can effectively improve accuracy. It can be understood that considering the cross entropy loss function, the gradient of the loss with respect to the log-odds is inversely proportional to the log-odds. When the gradient update of a sample is large, it indicates that the sample has a large amount of information, but it is more likely to be an outlier in the data distribution. In order to balance the noise and diversity in the training samples, this method can achieve this by adjusting the value of γ. A larger value of γ makes the small batch selector tend to extract samples with larger noise to enhance the training effect. Since the log-odds are usually small, the result is cubed to increase the difference in importance scores between different edges.

[0172] To be applied to Figure 3 As an example, after forward propagation, each positive sample e(u i ,v i ,t i )∈E B The importance weight w i =P(e) is updated according to the above function, and a batch of training edges E are randomly selected according to the corresponding importance score ratio B .

[0173] In the above embodiment, training samples are generated by adaptively small batches, the number of training times is reduced, the speed of model convergence is accelerated, and the training efficiency is improved, thereby improving the utilization rate of computer resources. In other words, this method can exclude some edge samples with general training effects or even side effects, thereby accelerating the convergence speed and reducing the training time.

[0174] Optionally, in an exemplary embodiment, the graph neural network is trained according to the structural position coding information and the common neighbor coding information, including: obtaining the original features of each node, the edge feature coding information and the time coding information of each edge; splicing the structural position coding information with the corresponding original features to obtain the node coding information of each node; splicing the node coding information, the edge feature coding information, the time coding information and the common neighbor coding information and mapping them to obtain the mapping features; using the mapping features to train the Transformer of the graph neural network. Among them, the original features can be understood as the node coding information, and the edge feature coding information and the time coding information can be implemented in the above manner, which will not be repeated here.

[0175] Optionally, in an exemplary embodiment, as Fig.10 As shown, a graph neural network training method is provided, which is applied to Figure 1 The terminal in is taken as an example to illustrate, including the following steps 1002 to 1020. Among them:

[0176] Step 1002, obtain the graph structure and the graph neural network to be trained.

[0177] The graph structure is determined based on the scheduled operation route information of the transportation tool, the nodes of the graph structure correspond to the target objects associated with the scheduled operation route information, the nodes are connected by edges, and the edges are used to represent the scheduled operation routes between the target objects.

[0178] Step 1004, if the current training of the graph neural network is the first training, all edge sets corresponding to the graph structure are divided based on the normal distribution to obtain multiple training edge sets.

[0179] Step 1006: Determine the neighbor position coding information of each neighbor node in the neighbor node set and the structural position coding information of each neighbor node in the graph structure.

[0180] Step 1008, determining the number of occurrences of each neighbor node in the N neighbor node sets, and expanding the common neighbors between the nodes according to the number of occurrences and the neighbor position coding information to obtain the common neighbor coding information.

[0181] Step 1010, training the graph neural network according to the structural position encoding information and the common neighbor encoding information to obtain a training result.

[0182] Step 1012: If the model has not converged, for each training from the start of the second training to the end of the i-th training, for each training edge set, determine the neighbor node set of each node in each edge.

[0183] Step 1014, update the weight of each edge in the graph structure according to the last training result to obtain the updated weight of each edge.

[0184] Step 1016 , extracting target edges from all edge sets corresponding to the graph structure according to the update weight of each edge, to obtain multiple training edge sets.

[0185] Step 1018, when the preset training conditions are met, the training is terminated to obtain a trained graph neural network.

[0186] Step 1020, obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the target operation route information in the graph structure to be processed.

[0187] It should be noted that the specific implementation method of this embodiment can be but is not limited to the above-mentioned limitation method, which will not be elaborated here.

[0188] In the above embodiment, by constructing position-differentiated coding and adding high-order neighbors, the information content of the input features is increased to enhance the model's perception of the graph structure and node positions. On this basis, from the perspective of reducing the single training time, it is proposed to fill the parallel sampler to reduce the sampling time, and optimize the common neighbor coding to reduce the time for encoding common neighbors. In addition, an adaptive small batch generation method based on the training results is used to accelerate convergence and improve training efficiency. In this way, a trained graph neural network is obtained, which in turn improves the accuracy of the model's prediction and improves the utilization rate of traffic resources in practical applications.

[0189] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0190] Based on the same inventive concept, the embodiment of the present application also provides a graph neural network training device for implementing the graph neural network training method involved above. The implementation solution provided by the device to solve the problem is similar to the implementation solution recorded in the above method, so the specific limitations in one or more graph neural network training device embodiments provided below can refer to the limitations on the graph neural network training method above, and will not be repeated here.

[0191] In an exemplary embodiment, Fig.11 As shown, a graph neural network training device is provided, including: a data acquisition module 1102, a neighbor node processing module 1104, a first encoding module 1106, a second encoding module 1108, a training module 1110 and a prediction module 1112, wherein:

[0192] The data acquisition module 1102 is used to obtain multiple training edge sets of a graph structure and a graph neural network to be trained. The graph structure is determined based on the scheduled operating route information of the transportation vehicle. The nodes of the graph structure correspond to target objects associated with the scheduled operating route information. The nodes are connected by edges, and the edges are used to represent the scheduled operating routes between the target objects.

[0193] The neighbor node processing module 1104 is used to determine the neighbor node set of each node in each edge for each training edge set.

[0194] The first encoding module 1106 is used to determine the neighbor position encoding information of each neighbor node in the neighbor node set and the structural position encoding information of each neighbor node in the graph structure.

[0195] The second encoding module 1108 is used to determine the number of occurrences of each neighbor node in the N neighbor node sets, and expand the common neighbors between the nodes according to the number of occurrences and the neighbor position encoding information to obtain common neighbor encoding information.

[0196] The training module 1110 is used to train the graph neural network according to the structural position coding information and the common neighbor coding information, and terminate the training to obtain a trained graph neural network when the preset training conditions are met.

[0197] The prediction module 1112 is used to obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

[0198] The above-mentioned graph neural network training device, through the graph structure, according to the characteristics of the graph structure and the dependencies between the nodes, splits the graph structure into multiple training edge sets for training the graph neural network, that is, uses small batches of training samples to train the graph neural network to be trained. On this basis, by considering the position of each edge node in the graph structure and the position of the node in the neighboring node in the training edge set, the corresponding neighbor position coding information and structure position coding information are obtained. On the basis of considering time coding, the neighbor position coding information and structure position coding information are added to enhance the information volume of the input features. In order to further improve the graph neural network's ability to represent the structure, by expanding the common neighbors between the nodes, more frequency domain information is captured, and common neighbor coding information is obtained. In this way, the information of the input features is enhanced to train the time-series graph neural network, solve the missing structural information in the model, and improve the representation ability of the time-series graph neural network for structural sensitivity, thereby obtaining a trained graph neural network. Based on the trained graph neural network, when processing the graph structure to be processed, the changes of the nodes over time and the time-sensitive characteristics can be better captured, and the rationality of the business information in the graph structure to be processed can be determined more comprehensively, thereby improving the accuracy and reliability of the prediction.

[0199] Optionally, in an exemplary embodiment, the business scenario includes at least one of a traffic prediction scenario, a course selection scenario, and a social network analysis scenario.

[0200] When the business scenario is a traffic prediction scenario, the graph structure is determined based on the scheduled operation route information of the transportation tool, the business node of the business scenario includes the target object associated with the scheduled operation route information, and the business information between the business nodes includes the scheduled operation route between the target objects;

[0201] In the scenario where the business scenario is a course selection scenario, the graph structure is determined based on the access behavior data of the selection object and the selected object. The business nodes of the business scenario include the selection object and the selected object, and the business information between the business nodes includes the access behavior between the selection object and the selected object.

[0202] When the business scenario is a social network analysis scenario, the graph structure is determined by the interactive behavior data in the social network, the business nodes of the business scenario include the target objects that generate the interactive behavior data, and the business information between the business nodes includes the interactive behaviors between the target objects.

[0203] Optionally, in an exemplary embodiment, the first encoding module includes a neighbor position encoding module and a structure position encoding module, wherein the neighbor position encoding module is used to sort each neighbor node in chronological order according to the timing information of each neighbor node in the neighbor node set, encode the sorted neighbor nodes, and obtain the neighbor position encoding information of each neighbor node.

[0204] The structural position encoding module is used to determine the number of paths between each neighbor node and the nodes, and determine the structural position encoding information of the neighbor node in the graph structure according to the number of paths.

[0205] Optionally, in an exemplary embodiment, the apparatus further comprises a sampling module, configured to sample neighbor nodes of each node in each edge in parallel for each training edge set, to obtain first-order neighbor nodes and high-order neighbor nodes of each node;

[0206] Optionally, in an exemplary embodiment, the neighbor node processing module 1104 is further configured to obtain a neighbor node set of each node according to the first-order neighbor nodes and the high-order neighbor nodes.

[0207] The second encoding module 1108 is further used to determine a common neighbor set of a first neighbor node set of a source node and a second neighbor node set of a target node;

[0208] For each neighbor node in the first neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding first two-dimensional vector;

[0209] For each neighbor node in the second neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding second two-dimensional vector;

[0210] Performing Fourier transform on the first two-dimensional vector and the second two-dimensional vector respectively to obtain first embedded representation and second embedded representation corresponding to each other;

[0211] According to the first embedding representation, the second embedding representation and the neighbor position encoding information, information of each common neighbor in the common neighbor set is determined to obtain common neighbor encoding information.

[0212] Optionally, in an exemplary embodiment, the apparatus further comprises a data acquisition module, configured to acquire original features of each node, edge feature coding information and time coding information of each edge;

[0213] Optionally, in an exemplary embodiment, the apparatus further comprises a splicing module, configured to splice the structural position coding information with the corresponding original features to obtain node coding information of each node;

[0214] Optionally, in an exemplary embodiment, the apparatus further comprises a mapping module, configured to concatenate and map the node coding information, the edge feature coding information, the time coding information and the common neighbor coding information to obtain a mapping feature;

[0215] Optionally, in an exemplary embodiment, the training module 1110 is also used to train the Transformer of the graph neural network using mapping features.

[0216] Optionally, in an exemplary embodiment, the apparatus further comprises a training sample generating module, the training sample generating module being configured to divide all edge sets corresponding to the graph structure based on a normal distribution to obtain a plurality of training edge sets if the current training of the graph neural network is the first training;

[0217] For each training from the start of the second training to the end of the i-th training, obtain the previous training result of the graph neural network;

[0218] Update the weight of each edge in the graph structure according to the last training result to obtain the updated weight of each edge;

[0219] According to the update weight of each edge, the target edge is extracted from all edge sets corresponding to the graph structure to obtain multiple training edge sets.

[0220] Optionally, in an exemplary embodiment, the above-mentioned device also includes a graph structure processing module, which is used to process a pre-stored graph structure to obtain a graph data structure that stores the connection relationship between nodes in the graph structure; and split the graph data structure to obtain all edge sets of the graph structure.

[0221] Each module in the above-mentioned graph neural network training device can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0222] In an exemplary embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Fig.12As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, near field communication (Near Field Communication, NFC) or other technologies. When the computer program is executed by the processor, a graph neural network training method is implemented. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0223] Those skilled in the art will understand that Fig.12 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0224] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0225] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0226] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0227] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0228] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0229] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0230] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A graph neural network training method, characterized in that: The method comprises: Acquire multiple training edge sets of a graph structure and a graph neural network to be trained, wherein the graph structure is determined according to a business scenario, the nodes of the graph structure correspond to business nodes of the business scenario, the nodes are connected by edges, and the edges are used to represent business information between the business nodes; For each of the training edge sets, determine a neighbor node set of each node in each edge; Determine neighbor position coding information of each neighbor node in the neighbor node set and structural position coding information of each neighbor node in the graph structure; Determine the number of occurrences of each of the neighbor nodes in the N neighbor node sets, and expand the common neighbors between the nodes according to the number of occurrences and the neighbor position encoding information to obtain common neighbor encoding information; The graph neural network is trained according to the structural position encoding information and the common neighbor encoding information, and when a preset training condition is met, the training is terminated to obtain a trained graph neural network; Obtain a graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain a prediction result of the graph structure to be processed.

2. The method according to claim 1, characterized in that The business scenario includes at least one of a traffic prediction scenario, a course selection scenario, and a social network analysis scenario. In the case where the business scenario is a traffic prediction scenario, the graph structure is determined according to the scheduled operation route information of the transportation tool, the business node of the business scenario includes the target object associated with the scheduled operation route information, and the business information between the business nodes includes the scheduled operation route between the target objects; In a scenario where the business scenario is a course selection scenario, the graph structure is determined according to access behavior data of a selection object and a selected object, the business node of the business scenario includes the selection object and the selected object, and the business information between the business nodes includes access behavior between the selection object and the selected object; In the case where the business scenario is a social network analysis scenario, the graph structure is determined by interactive behavior data in the social network, the business nodes of the business scenario include target objects that generate the interactive behavior data, and the business information between business nodes includes interactive behaviors between the target objects.

3. The method according to claim 1, characterized in that The determining of the neighbor position coding information of each neighbor node in the neighbor node set and the structural position coding information of each neighbor node in the graph structure includes: According to the timing information of each neighbor node in the neighbor node set, each of the neighbor nodes is sorted in chronological order, and the sorted neighbor nodes are encoded to obtain neighbor position encoding information of each neighbor node; Determine the number of paths between each of the neighboring nodes and the node; The structural position encoding information of the neighbor node in the graph structure is determined according to the number of paths.

4. The method according to claim 1, characterized in that: The step of determining, for each training edge set, a neighbor node set of each node in each edge includes: For each of the training edge sets, the neighbor nodes of each node in each edge are sampled in parallel to obtain the first-order neighbor nodes and high-order neighbor nodes of each node; A neighbor node set of each of the nodes is obtained according to the first-order neighbor nodes and the high-order neighbor nodes.

5. The method according to claim 1, characterized in that The nodes include a source node and a target node, and the determining the number of occurrences of each neighbor node in the N neighbor node sets, and expanding the common neighbors between the nodes according to the number of occurrences and neighbor position encoding information to obtain common neighbor encoding information, includes: Determine a common neighbor set of a first neighbor node set of the source node and a second neighbor node set of the target node; For each neighbor node in the first neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding first two-dimensional vector; For each neighbor node in the second neighbor node set of the source node, determine the number of times the neighbor node appears in the first neighbor node set and the second neighbor node set of the target node, and obtain a corresponding second two-dimensional vector; Performing Fourier transform on the first two-dimensional vector and the second two-dimensional vector respectively to obtain first embedded representation and second embedded representation corresponding to each other; According to the first embedded representation, the second embedded representation and the neighbor position encoding information, information of each common neighbor in the common neighbor set is determined to obtain common neighbor encoding information.

6. The method according to claim 1, characterized in that The training of the graph neural network according to the structural position encoding information and the common neighbor encoding information includes: Get the original features of each node, the edge feature encoding information and time encoding information of each edge; Concatenate the structural position coding information with the corresponding original features to obtain node coding information of each node; The node coding information, the edge feature coding information, the time coding information and the common neighbor coding information are concatenated and mapped to obtain a mapping feature; The mapping features are used to train the Transformer of the graph neural network.

7. The method according to any one of claims 1 to 6, characterized in that: The step of obtaining a plurality of training edge sets of a graph structure includes: If the current training of the graph neural network is the first training, all edge sets corresponding to the graph structure are divided based on normal distribution to obtain multiple training edge sets; For each training from the start of the second training to the end of the i-th training, obtaining the last training result of the graph neural network; The weight of each edge in the graph structure is updated according to the last training result to obtain an updated weight of each edge; According to the update weight of each of the edges, target edges are extracted from all edge sets corresponding to the graph structure to obtain multiple training edge sets.

8. The method according to claim 7, characterized in that The method further comprises: Processing a pre-stored graph structure to obtain a graph data structure storing connection relationships between nodes in the graph structure; The graph data structure is split to obtain all edge sets of the graph structure.

9. A graph neural network training device, characterized in that: The device comprises: A data acquisition module, used to acquire multiple training edge sets of a graph structure and a graph neural network to be trained, wherein the graph structure is determined according to a business scenario, the nodes of the graph structure correspond to business nodes of the business scenario, the nodes are connected by edges, and the edges are used to represent business information between the business nodes; A neighbor node processing module, used for determining, for each training edge set, a neighbor node set of each node in each edge; A first encoding module, used to determine neighbor position encoding information of each neighbor node in the neighbor node set and structural position encoding information of each neighbor node in the graph structure; A second encoding module is used to determine the number of occurrences of each of the neighbor nodes in the N neighbor node sets, and expand the common neighbors between the nodes according to the number of occurrences and the neighbor position encoding information to obtain common neighbor encoding information; A training module, used for training the graph neural network according to the structural position encoding information and the common neighbor encoding information, and terminating the training to obtain a trained graph neural network when a preset training condition is met; The prediction module is used to obtain the graph structure to be processed, input the graph structure to be processed into the trained graph neural network, and obtain the prediction result of the graph structure to be processed.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images