Dynamic network link prediction method based on high-order graph structure

Abstract

The invention provides a dynamic network link prediction method based on a high-order graph structure. The method comprises the following steps: extracting a plurality of high-order neighbor features of corresponding nodes through multi-hop sampling features; scoring the nodes, and filtering the nodes to obtain middle high-order neighbor features; constructing sub-graph network structure embedding, and cascading the middle high-order neighbor features through the sub-graph network structure embedding in combination with the evolution direction of the sub-graph network structure calculated by the network to obtain required high-order neighbor features; and performing feature fusion on the required high-order neighbor features to generate a target node embedded representation, and calculating a dynamic link prediction task based on the target node embedded representation. Through the method and the device, noise information with relatively low value can be filtered out, and the number of features is limited to a fixed size, so that the calculation complexity is reduced, the prediction accuracy of the model is improved, meanwhile, the memory occupation is optimized, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology, and in particular to a dynamic network link prediction method based on a high-order graph structure. Background Art

[0002] At present, when existing models use graph structure information to generate dynamic node embeddings, they only consider the one-hop neighbors or common neighbors of the nodes (i.e., first-order similarity and second-order similarity), resulting in insufficient expression of the network topology structure. In large-scale sparse networks, this limitation cannot fully describe the complex topological structure of the network and accurately measure the similarity between nodes, which limits the performance of the model in graph learning tasks such as node classification and link prediction.

[0003] In order to improve the expressiveness of network topology, high-order graph structures are introduced to capture the complex relationships and common features between nodes and provide a deeper similarity measure. However, their introduction is often accompanied by an exponential increase in computational complexity and a significant increase in memory usage, which is an obstacle that cannot be ignored in practical applications. In addition, high-order graph structures may also introduce unnecessary noise, affecting the final accuracy of the model. Therefore, further optimization is needed to solve these problems. Summary of the invention

[0004] Based on this, the purpose of the present invention is to provide a dynamic network link prediction method based on a high-order graph structure to solve the deficiencies in the prior art.

[0005] To achieve the above object, the present invention provides a dynamic network link prediction method based on a high-order graph structure, the method comprising: Through multi-hop sampling feature extraction, several high-order neighbor features of the corresponding node are obtained; Scoring the nodes by a preset node importance scoring function, filtering the nodes by a SortPooling layer based on the scores of the nodes, to obtain intermediate nodes and corresponding intermediate high-order neighbor features; Constructing a subgraph network structure embedding through the U-GCN embedding method, and cascading the intermediate high-order neighbor features through the subgraph network structure embedding and combining the evolution direction of the subgraph network structure calculated by the network to obtain the required high-order neighbor features; The required high-order neighbor features are optimized for feature fusion through Block-Recurrent Transformer to generate a target node embedding representation, and a dynamic link prediction task is calculated based on the target node embedding representation.

[0006] Preferably, the high-order neighbor features include time series features, neighbor co-occurrence features, node features and edge weight features, and the step of extracting several high-order neighbor features of the corresponding node through multi-hop sampling features includes: Collect N pieces of interaction data from the dynamic network link prediction dataset through multi-hop neighbors; A multi-hop neighbor set of a node is extracted from the N pieces of interaction data, and time series features, neighbor co-occurrence features, node features, and edge weight features are extracted from the multi-hop neighbor set, wherein corresponding One-Hot codes are respectively added to the tails of the node features and the edge weight features through a fully connected layer to distinguish neighbors with different hop numbers in the node features and the edge weight features.

[0007] Preferably, the expression of the node importance scoring function is as follows: The expression of the node importance scoring function is as follows:

[0008]

[0009] in, is the adjacency matrix of the sampled subgraph, is the identity matrix of the sampled sub-graphs, is the adjacency matrix and the identity matrix the sum of is the importance score of the node, Feature representation from extracted nodes, for The diagonal matrix of the matrix, For The matrix is ​​normalized. is a projection matrix with an output channel number of 1, is the activation function.

[0010] Preferably, the step of filtering the nodes through the SortPooling layer includes: The nodes are sorted from large to small based on the score value of each node, the first M nodes are retained, and the remaining nodes are filtered.

[0011] Preferably, the step of constructing a subgraph network structure embedding by the U-GCN embedding method comprises: Based on the intermediate nodes, sub-networks with different hop numbers of corresponding nodes are constructed, and information is propagated to each sub-network through the U-GCN network to obtain different network embedding matrices; Calculating the importance coefficient of the node in different network embedding matrices, and performing node neighbor aggregation on neighbors associated with the node based on the importance coefficient to obtain different network embeddings of the node; Different types of network embeddings are differentially aggregated to obtain subgraph network structure embeddings.

[0012] Preferably, before the step of cascading the intermediate high-order neighbor features, the method further comprises: Output the current subgraph network structure with the intermediate node as the core, read the hidden state of the subgraph network structure output last time from the subgraph structure embedding storage module, take the hidden state of the subgraph network structure output last time and the subgraph network structure embedding of the current subgraph network structure as input, calculate through the network, and obtain the evolution direction of the current subgraph network structure.

[0013] Preferably, the step of cascading the intermediate high-order neighbor features to obtain the required high-order neighbor features comprises: Based on the subgraph network structure embedding and the evolution direction, multiple subnetworks are generated through dynamic node embedding, and the features output by the first two subnetworks are sequentially blocked, aligned and cascaded to generate the required high-order neighbor features of the relevant nodes.

[0014] Preferably, the step of performing feature fusion optimization on the required high-order neighbor features through Block-Recurrent Transformer comprises: The Block-Recurrent Transformer is stacked vertically by L-1 layers and horizontally by 1 layer to obtain the target Block-Recurrent Transformer, where L is the number of layers; Dividing the required high-order neighbor features into B blocks through the target Block-Recurrent Transformer, and embedding the B blocks as input tokens; Through the attention mechanism and in combination with the current state vector, each block is cross-attended with a recurrent unit so that the elements of each block are attended to the summary of the elements in the previous block to obtain the current output of each layer in the target Block-Recurrent Transformer, wherein the current output includes the current output token embedding and the next state vector, and the current output token embedding is represented as the target node embedding.

[0015] Preferably, the step of calculating the dynamic link prediction task based on the target node embedding representation comprises: The hidden representation of the current output token embedding is pooled and outputted through the pooling layer to obtain the final node embedding vector; The node embedding vector is calculated through the MLP network to obtain an output dynamic link prediction result.

[0016] The beneficial effects of the present invention are as follows: a plurality of high-order neighbor features of the corresponding node are obtained through multi-hop sampling feature extraction, the node is scored through a preset node importance scoring function, based on the score of each node, the node is filtered through a SortPooling layer to obtain intermediate nodes and corresponding intermediate high-order neighbor features, and then a subgraph network structure embedding is constructed through a U-GCN embedding method, and the intermediate high-order neighbor features are cascaded through the subgraph network structure embedding and the evolution direction of the subgraph network structure to obtain the required high-order neighbor features to enhance the expression ability of the network topology structure, and the required high-order neighbor features are feature fused through a Block-Recurrent Transformer to generate a target node embedding representation, and then a dynamic link prediction task is calculated based on the target node embedding representation. Different from the prior art, higher-order feature extraction of node dimensions can be realized to provide more available information, and noise information with lower value can be filtered out, the number of features is limited to a fixed size, the complexity of calculation is reduced, and the prediction accuracy of the model is improved. At the same time, the memory usage is optimized and the calculation efficiency is improved.

[0017] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of a dynamic network link prediction method based on a high-order graph structure provided by an embodiment of the present invention; Figure 2 A framework diagram of a continuous time attention network model based on subgraph information aggregation and network structure embedding provided by an embodiment of the present invention; Figure 3 A framework diagram of a subgraph structure embedding module provided in an embodiment of the present invention; Figure 4 A framework diagram of a subgraph structure evolution module provided in an embodiment of the present invention; Figure 5 A framework diagram of the Block-Recurrent Transformer module provided in an embodiment of the present invention.

[0019] The following specific implementation manner will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated 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 intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0021] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. In addition, it can also be understood that although the efforts made in this development process may be complicated and lengthy, for ordinary technicians in this field related to the content disclosed in this application, some changes in design, manufacturing or production based on the technical content disclosed in this application are just conventional technical means, and should not be understood as insufficient content disclosed in this application.

[0022] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0023] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantity limitation, and may indicate the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0024] See also Figure 1 , which is a flow chart of a dynamic network link prediction method based on a high-order graph structure in an embodiment of the present invention. This method proposes a continuous-time attention network model based on subgraph information aggregation and network structure embedding from three perspectives: multi-hop sampling feature extraction, subgraph network structure embedding, and feature fusion optimization. This model is used to implement a dynamic network link prediction method based on a high-order graph structure. Compared with similar algorithms, this model can achieve better dynamic network link prediction effects.

[0025] The model is mainly divided into three parts, namely, the multi-hop feature extraction and filtering sub-network, the sub-graph network structure embedding sub-network and the dynamic node embedding generation sub-network. The multi-hop feature extraction and filtering sub-network includes the multi-hop historical interaction extraction module, the feature extraction module and the node sorting and filtering module. The sub-graph network structure embedding sub-network includes the sub-graph structure embedding module, the sub-graph structure evolution module and the sub-graph structure embedding storage module. The dynamic node embedding generation sub-network includes the Block-Recurrent Transformer module. The basic framework of the model is as follows: Figure 2 As shown in the figure, in this model, multi-hop sampling features are extracted, the subgraph size is limited by combining the score sorting and filtering mechanism, and the subgraph structure embedding is constructed. Finally, the Block-Recurrent Transformer is used for feature fusion to generate node representations for the task of dynamic link prediction.

[0026] like Figure 1 The dynamic network link prediction method based on high-order graph structure shown in the figure comprises the following steps: Step S101, extracting several high-order neighbor features of the corresponding node through multi-hop sampling features; Among them, through the multi-hop sampling feature extraction method, more high-order features can be obtained, providing more available information to improve the final prediction accuracy of the continuous-time attention network model based on sub-graph information aggregation and network structure embedding.

[0027] Step S102, scoring the nodes by a preset node importance scoring function, filtering the nodes by a SortPooling layer based on the scores of the nodes, to obtain intermediate nodes and corresponding intermediate high-order neighbor features; Among them, this step is implemented through the node sorting and filtering module. The node importance scoring function can be customized. The goal of this module is to limit the size of the subgraph passed in by the subsequent subgraph structure embedding module to a fixed size while controlling the computational complexity, and filter out some noise information with low value in the early multi-hop sampling. The specific implementation is, first, score each node using the node importance scoring function, and then, according to the score value of each node, retain the most important top M node information through the SortPooling layer and pass it to the subsequent sub-network for processing.

[0028] Finally, the node sorting and filtering module will output the nodes after importance sorting and filtering. and Related features and , which contain nodes and The sub-graph network structure with the core and various features extracted by the feature extraction module.

[0029] Step S103, constructing a subgraph network structure embedding through the U-GCN embedding method, and cascading the intermediate high-order neighbor features through the subgraph network structure embedding and combining the evolution direction of the subgraph network structure calculated by the network to obtain the required high-order neighbor features; Among them, the main idea of ​​the U-GCN embedding method is to use the relevant features passed by the node sorting and filtering module and , respectively construct the relevant nodes and The subgraph network structure embedding and the evolution direction of the subgraph network structure.

[0030] Step S104, performing feature fusion optimization on the required high-order neighbor features through Block-Recurrent Transformer to generate a target node embedding representation, and calculating a dynamic link prediction task based on the target node embedding representation.

[0031] Among them, since the computational cost of the attention layer in the Block-Recurrent Transformer increases linearly with the sequence length, the Block-Recurrent Transformer has a lower computational complexity for the step of feature fusion optimization that requires high-order neighbor features, thereby optimizing memory usage.

[0032] Through the above steps, several high-order neighbor features of the corresponding node are obtained by multi-hop sampling feature extraction, the nodes are scored by a preset node importance scoring function, based on the scores of each node, the nodes are filtered through the SortPooling layer to obtain intermediate nodes and corresponding intermediate high-order neighbor features, and then the subgraph network structure embedding is constructed by the U-GCN embedding method. Through the subgraph network structure embedding and the evolution direction of the subgraph network structure, the intermediate high-order neighbor features are cascaded to obtain the required high-order neighbor features to enhance the expression ability of the network topology structure, and the required high-order neighbor features are feature fused by Block-Recurrent Transformer to generate the target node embedding representation, and then the dynamic link prediction task is calculated based on the target node embedding representation. Different from the existing technology, it can realize higher-order feature extraction of node dimensions to provide more available information, which is conducive to improving the accuracy of the final prediction, and can also filter out noise information with lower value, limit the number of features to a fixed size, reduce the complexity of calculation, and at the same time, optimize memory usage and improve calculation efficiency.

[0033] In some embodiments, the high-order neighbor features include time series features, neighbor co-occurrence features, node features, and edge weight features, and the step of extracting several high-order neighbor features of the corresponding node through multi-hop sampling features includes: Collect N pieces of interaction data from the dynamic network link prediction dataset through multi-hop neighbors; This step is implemented by the multi-hop history interaction extraction module, which mainly expands the previous one-hop neighbor sampling to Hop-neighbor sampling.

[0034] A multi-hop neighbor set of a node is extracted from the N pieces of interaction data, and time series features, neighbor co-occurrence features, node features, and edge weight features are extracted from the multi-hop neighbor set, wherein corresponding One-Hot codes are respectively added to the tails of the node features and the edge weight features through a fully connected layer to distinguish neighbors with different hop numbers in the node features and the edge weight features.

[0035] Among them, this step is implemented by the feature extraction module, which is mainly The collection after skipping sampling and The elements in are used to extract time series features, node features, edge weight features, and neighbor co-occurrence features. On the basis of encoding the node features and edge weight features using the fully connected layer, an additional One-Hot encoding is added at the end to distinguish neighbors with different hop counts. For example, [1, 0, 0] is used to encode one-hop neighbors, [0, 1, 0] is used to encode two-hop neighbors, and [0, 0, 1] is used to encode three-hop neighbors.

[0036] In some embodiments, the expression of the node importance scoring function is as follows: The expression of the node importance scoring function is as follows:

[0037]

[0038] in, is the adjacency matrix of the sampled subgraph, is the identity matrix of the sampled sub-graphs, is the adjacency matrix and the identity matrix the sum of is the importance score of the node, Feature representation from extracted nodes, Features from extracted nodes u and v and , for The diagonal matrix of the matrix, For The matrix is ​​normalized. is a projection matrix with an output channel number of 1, is the activation function.

[0039] In some embodiments, the step of filtering the nodes through the SortPooling layer includes: The nodes are sorted from large to small based on the score value of each node, the first M nodes are retained, and the remaining nodes are filtered.

[0040] Among them, this step is implemented through a node sorting and filtering module, specifically, the importance of the corresponding nodes is sorted according to the scoring value of the node, and the importance is sorted from large to small. The first M relatively more important nodes are retained, and the remaining nodes are filtered, and the filtered nodes are regarded as noise information.

[0041] In some embodiments, the step of constructing a subgraph network structure embedding by the U-GCN embedding method includes: Based on the intermediate nodes, sub-networks with different hop numbers of corresponding nodes are constructed, and information is propagated to each sub-network through the U-GCN network to obtain different network embedding matrices; Calculating the importance coefficient of the node in different network embedding matrices, and performing node neighbor aggregation on neighbors associated with the node based on the importance coefficient to obtain different network embeddings of the node; Different types of network embeddings are differentially aggregated to obtain subgraph network structure embeddings.

[0042] The above steps are implemented by the subgraph structure embedding module. Specifically, as for the implementation of the subgraph structure embedding module, Figure 3 As shown, the node For example, for the input subgraph , respectively constructed with nodes One-hop network , two-hop network and kNN network ,in, , and Represent the adjacency matrices of the three networks respectively. Represent the node feature matrix respectively. After the GCN network carries out information propagation process, three network embedding matrices are obtained , and . Then we need to calculate the node pair After obtaining the importance coefficients between node pairs in the three network types, we will then The associated neighbors perform node neighbor aggregation to obtain the relevant nodes Three network embeddings , and Finally, perform differential aggregation on the network type to obtain the final subgraph network structure embedding , for nodes By following the same process as above, we can get the final subgraph network structure embedding .

[0043] In some embodiments, before the step of cascading the intermediate high-order neighbor features, the method further includes: Output the current subgraph network structure with the intermediate node as the core, read the hidden state of the subgraph network structure output last time from the subgraph structure embedding storage module, take the hidden state of the subgraph network structure output last time and the subgraph network structure embedding of the current subgraph network structure as input, calculate through the network, and obtain the evolution direction of the current subgraph network structure.

[0044] This step is implemented through the subgraph structure evolution module. The main contents of the subgraph structure evolution module are as follows: Figure 4 First, the subgraph structure embedding module outputs a node-based and The core subgraph network structure is embedded. Then, through the node and The id of each subgraph is used to read the hidden state of the subgraph network structure output last time from the subgraph structure embedding storage module. , Next, embed the hidden state of the subgraph network structure output last time and the subgraph network structure of the current subgraph network structure output this time into , As input, after the gated recurrent unit network calculation, the hidden state of the current subgraph network structure is obtained , The evolution direction of the current subgraph network structure is represented by the difference between the hidden state of the current subgraph network structure and the hidden state of the subgraph network structure output last time.

[0045] In some embodiments, the step of cascading the intermediate high-order neighbor features to obtain the required high-order neighbor features includes: Based on the subgraph network structure embedding and the evolution direction, multiple subnetworks are generated through dynamic node embedding, and the features output by the first two subnetworks are sequentially blocked, aligned and cascaded to generate the required high-order neighbor features of the relevant nodes.

[0046] First, the nodes transferred from the multi-hop feature extraction filtering subnetwork are used and Corresponding node neighbor features and , nodes passed from the subgraph network structure embedded in the subnetwork and The corresponding subgraph network structure is embedded and and nodes and Corresponding subgraph network structure evolution direction and , get the cascaded features , and then sent to the subsequent Block-Recurrent Transformer module, For Node The corresponding cascaded features, For Node The corresponding cascaded features.

[0047] In some embodiments, the step of performing feature fusion optimization on the required high-order neighbor features through Block-Recurrent Transformer includes: The Block-Recurrent Transformer is stacked vertically by L-1 layers and horizontally by 1 layer to obtain the target Block-Recurrent Transformer, where L is the number of layers; Dividing the required high-order neighbor features into B blocks through the target Block-Recurrent Transformer, and embedding the B blocks as input tokens; Through the attention mechanism and in combination with the current state vector, each block is cross-attended with a recurrent unit so that the elements of each block are attended to the summary of the elements in the previous block to obtain the current output of each layer in the target Block-Recurrent Transformer, wherein the current output includes the current output token embedding and the next state vector, and the current output token embedding is represented as the target node embedding.

[0048] The above steps are implemented by the Block-Recurrent Transformer module. The Block-Recurrent Transformer module mainly uses the Block-Recurrent Transformer to cascade the features. Perform feature fusion optimization to obtain the fused feature representation , and then the node is obtained through the pooling output layer and The corresponding dynamic node representation and , For Node The corresponding fused features are For Node The corresponding fused features.

[0049] In terms of specific implementation, Figure 5 As shown, the Block-Recurrent Transformer module will pass the features , divided into B blocks, and embedded as input tokens. Block division effectively expands the amount of information that can be processed, so that a larger attention window can be utilized. On each individual block, by applying local attention and combining a set of current state vectors, each block outputs the summary of its elements to the elements of the next block after passing through the recurrent network, realizing cross-attention between two adjacent blocks, so that the elements in one block can pay attention to the summary of the elements in the previous block. The cross-block attention mechanism can better capture the long-term dependencies in the sequence. In the model implementation setting, this paper vertically stacks the Block-Recurrent Transformer with a total of L-1 layers, and then horizontally stacks 1 layer to fully learn The internal dependencies of the feature sequence, each layer output includes the generated token embedding and the next state vector.

[0050] In some embodiments, the step of calculating the dynamic link prediction task based on the target node embedding representation includes: The hidden representation of the current output token embedding is pooled and outputted through the pooling layer to obtain the final node embedding vector; The node embedding vector is calculated through the MLP network to obtain an output dynamic link prediction result.

[0051] The output of each block is concatenated into a matrix ,matrix The dimensions and input matrix Then, the dynamic node representation is obtained by using a pooling method based on mean pooling. and , the dynamic node represents the dynamic link prediction result used to calculate.

[0052] It should be noted that the above implementation process is only to illustrate the feasibility of the present application, but this does not mean that the dynamic network link prediction method based on the high-order graph structure of the present application has only the above-mentioned single implementation process. On the contrary, as long as the dynamic network link prediction method based on the high-order graph structure of the present application can be implemented, it can be included in the feasible implementation plan of the present application.

[0053] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0054] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. 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 invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A dynamic network link prediction method based on high-order graph structure, characterized in that: The method comprises: Through multi-hop sampling feature extraction, several high-order neighbor features of the corresponding node are obtained; Scoring the nodes by a preset node importance scoring function, filtering the nodes by a SortPooling layer based on the scores of the nodes, to obtain intermediate nodes and corresponding intermediate high-order neighbor features; Constructing a subgraph network structure embedding through the U-GCN embedding method, and cascading the intermediate high-order neighbor features through the subgraph network structure embedding and combining the evolution direction of the subgraph network structure calculated by the network to obtain the required high-order neighbor features; The required high-order neighbor features are optimized for feature fusion through Block-Recurrent Transformer to generate a target node embedding representation, and a dynamic link prediction task is calculated based on the target node embedding representation.

2. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: The high-order neighbor features include time series features, neighbor co-occurrence features, node features, and edge weight features. The step of extracting several high-order neighbor features of the corresponding node through multi-hop sampling features includes: Collect N pieces of interaction data from the dynamic network link prediction dataset through multi-hop neighbors; A multi-hop neighbor set of a node is extracted from the N pieces of interaction data, and time series features, neighbor co-occurrence features, node features, and edge weight features are extracted from the multi-hop neighbor set, wherein corresponding One-Hot codes are respectively added to the tails of the node features and the edge weight features through a fully connected layer to distinguish neighbors with different hop numbers in the node features and the edge weight features.

3. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: The expression of the node importance scoring function is as follows: in, is the adjacency matrix of the sampled subgraph, is the identity matrix of the sampled sub-graphs, is the adjacency matrix and the identity matrix the sum of is the importance score of the node, Feature representation from extracted nodes, for The diagonal matrix of the matrix, For The matrix is ​​normalized. is a projection matrix with an output channel number of 1, is the activation function.

4. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: The step of filtering the nodes through the SortPooling layer includes: The nodes are sorted from large to small based on the score value of each node, the first M nodes are retained, and the remaining nodes are filtered.

5. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: The step of constructing a subgraph network structure embedding by the U-GCN embedding method comprises: Based on the intermediate nodes, sub-networks with different hop numbers of corresponding nodes are constructed, and information is propagated to each sub-network through the U-GCN network to obtain different network embedding matrices; Calculating the importance coefficient of the node in different network embedding matrices, and performing node neighbor aggregation on neighbors associated with the node based on the importance coefficient to obtain different network embeddings of the node; Different types of network embeddings are differentially aggregated to obtain subgraph network structure embeddings.

6. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: Before the step of cascading the intermediate high-order neighbor features, the method further includes: Output the current subgraph network structure with the intermediate node as the core, read the hidden state of the subgraph network structure output last time from the subgraph structure embedding storage module, take the hidden state of the subgraph network structure output last time and the subgraph network structure embedding of the current subgraph network structure as input, calculate through the network, and obtain the evolution direction of the current subgraph network structure.

7. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: The step of cascading the intermediate high-order neighbor features to obtain the required high-order neighbor features comprises: Based on the subgraph network structure embedding and the evolution direction, multiple subnetworks are generated through dynamic node embedding, and the features output by the first two subnetworks are sequentially blocked, aligned and cascaded to generate the required high-order neighbor features of the relevant nodes.

8. The dynamic network link prediction method based on high-order graph structure according to claim 1 is characterized in that: The step of performing feature fusion optimization on the required high-order neighbor features through Block-Recurrent Transformer includes: The Block-Recurrent Transformer is stacked vertically by L-1 layers and horizontally by 1 layer to obtain the target Block-Recurrent Transformer, where L is the number of layers; Dividing the required high-order neighbor features into B blocks through the target Block-Recurrent Transformer, and embedding the B blocks as input tokens; Through the attention mechanism and in combination with the current state vector, each block is cross-attended with a recurrent unit so that the elements of each block are attended to the summary of the elements in the previous block to obtain the current output of each layer in the target Block-Recurrent Transformer, wherein the current output includes the current output token embedding and the next state vector, and the current output token embedding is represented as the target node embedding.

9. The method for dynamic network link prediction based on high-order graph structure according to claim 8, characterized in that: The step of calculating the dynamic link prediction task based on the target node embedding representation comprises: The hidden representation of the current output token embedding is pooled and outputted through the pooling layer to obtain the final node embedding vector; The node embedding vector is calculated through the MLP network to obtain an output dynamic link prediction result.

Images