A Dynamic Link Prediction Method and System Based on Multi-Granularity Evolution

Abstract

The present invention relates to a dynamic link prediction method and system based on multi-granularity evolution. The method includes: dividing the graph under each time slice of the dynamic graph to obtain multi-granularity subgraphs; extracting the structural features of nodes from the multi-granularity subgraphs; learning the dynamic evolution law of the graph from the multi-granularity subgraphs to obtain the dynamic evolution features of subgraphs with different granularities, and fusing the dynamic evolution features of subgraphs with different granularities to obtain the dynamic evolution features of nodes; fusing the structural features of nodes with the dynamic evolution features of nodes to obtain node representations containing spatio-temporal features, and predicting future links according to the node representations containing spatio-temporal features. The present invention can mine richer graph information, thereby enhancing the ability to extract structural features and improving the accuracy of link prediction; the present invention can fully realize the fusion of dynamic features, enhance the ability to learn dynamic features, and thus improve the performance of link prediction.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and particularly relates to a dynamic link prediction method and system based on multi-granularity evolution. Background Art

[0002] Many complex information systems in the real world can be modeled in the form of a graph, and in recent years, the research on graph data has been continuously deepening. Since the relationship information between entities is crucial for the association calculation on the graph, and link prediction can predict the possibility of the existence of a relationship between unformed nodes, thereby mining effective information, it is one of the important tasks in the field of graph mining. At the same time, the graph data in the real world often evolves dynamically over time. Link prediction for dynamic graphs is more valuable. Its goal is to learn the structural evolution of the graph from past information to predict future links, which has attracted extensive attention. For example, Chinese Patent (Application No.: 201910440098.5, Publication No.: CN 110413844 A) proposes a dynamic link prediction method based on a spatio-temporal attention deep model. In the encoding stage, this method takes the sequence of adjacency matrices of different time slices of the dynamic graph as input. In the GCN-attention module, a spatial attention mechanism is used to focus on neighbor nodes to update the feature vectors of each node under the corresponding time slice. Then, the feature vectors of T time slices are input into the LSTM-attention module to obtain hidden layer vectors, and the context vectors are calculated for the hidden layer vectors of T times. The context vector is used as the spatio-temporal feature vector and input into the decoder. The decoder decodes the input spatio-temporal feature vector to output a probability matrix representing whether there is a link between nodes, that is, the prediction of the dynamic link is realized. Chinese Patent (Application No.: 202110280461.9, Publication No.: CN 113065974 A) proposes a link prediction method based on dynamic network representation learning. This method constructs a similarity matrix of the snapshot network by calculating the similarity values between nodes of the dynamic network through a similarity-based aggregation strategy; applies a graph convolutional neural network to a single snapshot network for feature aggregation, and uses the adjacency matrix and the similarity matrix to guide the feature aggregation process to determine the low-dimensional feature representation of the nodes; inputs the low-dimensional feature representation of the nodes into a logistic regression classifier to obtain the link prediction result of the dynamic network.

[0003] Existing dynamic link prediction methods focus on extracting the structural and dynamic features of dynamic graphs across the entire graph. However, these methods ignore that graphs at different levels contain different structures and change patterns, resulting in imperfect feature extraction and inaccurate dynamic link prediction. For example, in a social network, there are different groups, such as student groups, corporate colleague groups, etc. Interactions (sending messages, emails) between different groups over a period of time form a dynamic graph. Some groups have more interactions and dense structures, while some groups have fewer interactions and relatively sparse structures. Subgraphs formed by different groups contain different structures and dynamic change patterns. Learning the structural and dynamic features of different groups discriminatively is beneficial for mining richer dynamic graph information. If features are extracted across the entire graph, it may introduce some noise, resulting in fewer predicted edges than actual in dense structures and more predicted edges than actual in sparse structures, and the overall dynamic link prediction effect is poor. Summary of the Invention

[0004] Aiming at the deficiencies of existing link prediction methods, the present invention proposes a dynamic link prediction method and system based on multi-granularity evolution.

[0005] The present invention first divides the graph under each time slice according to a certain strategy to construct multi-granularity subgraphs. Then, a multi-granularity-based structural feature extraction module is designed to extract and fuse features from subgraphs at different granularities, thereby capturing the structural features of the graph. Next, a multi-granularity-based dynamic evolution fusion module is designed to learn the dynamic laws of the graph from subgraphs at different granularities. Further, by considering the different importance of features at different granularities, these dynamic features are selectively fused and the structural features are incorporated at the same time, finally obtaining a node representation containing spatio-temporal features, so as to more accurately predict the future relationships between entities.

[0006] Specifically, the technical solutions adopted by the present invention are as follows:

[0007] A dynamic link prediction method based on multi-granularity evolution, comprising the following steps:

[0008] Divide the graph under each time slice of the dynamic graph to obtain multi-granularity subgraphs;

[0009] Extract the structural features of nodes from the multi-granularity subgraphs;

[0010] Learn the dynamic evolution laws of the graph from the multi-granularity subgraphs, obtain the dynamic evolution features of subgraphs at different granularities, and fuse the dynamic evolution features of subgraphs at different granularities to obtain the dynamic evolution features of nodes;

[0011] Fuse the structural features of nodes with the dynamic evolution features of nodes to obtain a node representation containing spatio-temporal features, and predict future links according to the node representation containing spatio-temporal features.

[0012] Further, the k-truss subgraph decomposition algorithm is used to partition the graph under each time slice to obtain the multi-granularity subgraphs.

[0013] Further, a graph convolutional network based on disentangled propagation and mapping operations is used and applied to the k-truss subgraphs to extract the structural features of the nodes.

[0014] Further, the formula of the graph convolutional network based on disentangled propagation and mapping operations is as follows:

[0015]

[0016] att l = σ(MLP(X l )) l = 0, 1, 2, …, J

[0017] X out = softmax(sum(att 0 ο X 0 , …, att J ο X J ))

[0018] where l is the number of GCN layers; J is the maximum depth of propagation; is the node feature, N represents the number of nodes, d represents the feature dimension; MLP represents the fully connected operation; X l represents the node feature at depth l; is used to adaptively adjust the information that each node needs to retain at different propagation depths; ° represents that each element in att l is multiplied by the corresponding d-dimensional node representation X J ; softmax represents the normalization operation; sum represents the summation operation; the final output X out is obtained by combining multiple propagation layers.

[0019] Further, learning the dynamic evolution law of the graph from the multi-granularity subgraphs means using an RNN to learn the dynamic evolution laws of subgraphs with different granularities.

[0020] Further, fusing the dynamic evolution features of subgraphs with different granularities to obtain the dynamic evolution features of nodes includes: introducing a multi-head self-attention mechanism, learning the weights on each granularity through the self-attention mechanism, and then multiplying the weights by the node representations of the corresponding granularities and summing them to obtain the fused representation of the dynamic evolution features of nodes.

[0021] Furthermore, predicting future links based on node representations containing spatiotemporal features includes: performing a Hadamard product operation on two node representations corresponding to the link, predicting the link score through the node representation, and then training a logistic regression classifier to determine whether the link exists, thereby predicting the link on the graph at the next time.

[0022] A dynamic link prediction system based on multi-granularity evolution using the above method comprises:

[0023] The multi-granularity graph partitioning module is used to partition the graph in each time slice of the dynamic graph to obtain multi-granularity subgraphs;

[0024] Multi-granularity structural feature extraction module, used to extract the structural features of nodes from multi-granularity subgraphs;

[0025] The multi-granularity dynamic feature learning module is used to learn the dynamic evolution law of the graph from the multi-granularity subgraph, obtain the dynamic evolution features of the subgraphs of different granularities, and fuse the dynamic evolution features of the subgraphs of different granularities to obtain the dynamic evolution features of the nodes;

[0026] The dynamic link prediction module is used to fuse the structural features of the node with the dynamic evolution features of the node to obtain a node representation containing spatiotemporal features, and predict future links based on the node representation containing spatiotemporal features.

[0027] The beneficial effects of the present invention are as follows:

[0028] 1. In view of the fact that existing dynamic link prediction methods focus on extracting features on the entire graph and ignore the problem that subgraph structures at different granularities have different structural features, the present invention divides the graph into multi-granularity subgraphs according to a certain strategy, and then extracts features from each subgraph. The subgraphs of different granularities in the same time slice are linked together to learn their development pattern from a locally dense structure to a globally sparse structure, thereby mining richer graph information, thereby enhancing the ability to extract structural features and improving the accuracy of link prediction.

[0029] 2. Aiming at the problem of insufficient dynamic feature learning in dynamic link prediction methods, the present invention proposes a method for multi-granularity dynamic feature evolution fusion, which learns the dynamic evolution features of subgraph time series at different granularities and obtains dynamic features of multiple granularities. Then, combined with the self-attention mechanism, different weights are set for features of different granularities to increase the contribution of important features, fully realize the fusion of dynamic features, enhance the learning ability of dynamic features, and thus improve the performance of link prediction.

[0030] 3. The present invention can be used in fields such as social networks and recommendation systems. For example, when used in a social network, the nodes in the present invention can represent users in the social network, and the predicted links can represent potential (future) friendship relationships between users in the social network. When used in a recommendation system, the nodes in the present invention can represent users and commodities in the recommendation system. The predicted links connect users and commodities, and can represent recommending potential commodities to users (potential commodities refer to commodities that users like but have not purchased yet). BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is an example of a k-truss graph.

[0032] Figure 2 is a schematic flow chart of the method of the present invention.

[0033] Figure 3 is a schematic diagram of the dynamic link prediction architecture proposed by the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The present invention will be further described in detail below through specific embodiments and the accompanying drawings.

[0035] The dynamic link prediction method based on multi-granularity evolution proposed by the present invention provides a complete algorithm framework, which mainly consists of four parts: construction of a dynamic multi-granularity graph, extraction of multi-granularity structural features, evolution fusion of multi-granularity dynamic features, and dynamic link prediction. The specific steps are as follows:

[0036] Step 1. Construction of a dynamic multi-granularity graph

[0037] The present invention defines the dynamic graph G as a time series sequence {G1, G2, …, G T}, where the time t ranges from 1 to T. A certain strategy is needed to separate the diverse structures of the graph at each time, and rich discriminative structural information is mined to help improve the ability of link prediction. According to the requirements, the present invention introduces the k-truss subgraph decomposition algorithm in the field of community mining. First, the concepts of support and k-truss subgraph are introduced and defined as follows:

[0038] Definition 1. Support(e): If an edge e in the graph G is in k triangles, then the support of the edge support(e) = k - 2.

[0039] Definition 2. k-truss: Given a graph G and an integer k, the k-truss subgraph is a maximal subgraph g in G. In the maximal subgraph g, the support of all edges support(e) >= k - 2.

[0040] Figure 1Shows the subgraphs decomposed according to the k-truss definition. Among them, the leftmost one is the original graph and remains unchanged. According to the definition, the edge support in the 2-truss subgraph should be greater than 0, which means that each edge should be in at least 0 triangles. Any edge that satisfies this condition is the graph itself. By analogy, each edge in the 3-truss subgraph is in at least 1 triangle, as shown in the second figure on the right. The 4-truss subgraph is shown in the rightmost figure.

[0041] Thus, multi-granularity subgraphs divided by k-truss decomposition are obtained. The multi-granularity subgraphs obtained by the k-truss strategy have many excellent characteristics. The k-truss subgraph requires that all edges be in at least (k - 2) triangles. As we all know, triangles are the basic modules of real networks because a link is caused by triangle closure, and triangle closure has a high local clustering coefficient. In a social network, a triangle means that two friends have a common friend, indicating a strong and stable relationship among three friends. Further thinking, if two people have more friends, then their friendship is more stable, which can emphasize the reliability of the relationship between any two friends. In the link prediction task, the k-truss strategy considers the common neighbors of nodes. Subgraphs with higher k values have a denser structure and fewer nodes. Intuitively, the more common neighbors two nodes have, the more likely they are to become friends or interact, so it has strong advantages.

[0042] Apply the k-truss decomposition algorithm to the graph at each time slice to obtain the multi-granularity subgraph sequence corresponding to that time slice.

[0043] Step 2. Multi-granularity structure feature learning.

[0044] To extract sufficient features from the multi-granularity graphs obtained from multiple time slices to reflect the structural diversity, we need to propagate features in a larger domain. One of the most commonly used methods is GCN (Graph Convolution Network), which extracts features through two operations: propagation and mapping. The most classic GCN formula is as follows:

[0045]

[0046] Among them, X l is the node representation matrix, and the adjacency matrix with self-loops is the adjacency matrix A with the identity matrix I added. Each element a ij in the adjacency matrix represents whether there is an edge between node i and node j. 0 represents no edge, and 1 represents an edge. is the degree matrix with self-loops, where diag() represents the diagonal matrix, and the elements on the diagonal represent the degrees of the corresponding nodes. Represents The value at the \(i\)-th row and \(j\)-th column, which indicates whether there is an edge between node \(i\) and node \(j\). The superscript \(l\) represents the layer number of the GCN, and \(W\) l Represents the parameters of the \(l\)-th layer, and \(\sigma()\) is a non-linear activation function. Here, the propagation operation is Used to transfer node features to neighboring nodes. The mapping operation is \(\sigma(P l-1 W l ), which is used to map the node representation to the feature space.

[0047] However, this way of entangle propagation and mapping operations leads to limited propagation depth of the GCN and cannot propagate features in large domains. Therefore, the present invention designs a graph convolutional network based on disentangled propagation and mapping operations, named Modified GCN, to extract features, which can alleviate the overfitting problem of general GCN and thus propagate information in a larger range. Its formula is as follows:

[0048]

[0049] att l =\(\sigma(MLP(X l ))\) \(l = 0, 1, 2, \ldots, J\)

[0050] X out = softmax(\(\sum\)(att 0 \(\circ\)X 0 , \ldots, att J \(\circ\)X J ))

[0051] where \(l\) is the layer number of the GCN and also the depth. \(J\) is the maximum depth of propagation, is the node feature, \(N\) represents the number of nodes, \(d\) represents the feature dimension, and MLP represents the fully connected operation. \(X l represents the node feature at depth \(l\). Used to adaptively adjust the information that each node needs to retain at different propagation depths. \(\circ\) represents that each element in att l is multiplied by the corresponding \(d\)-dimensional node representation \(X J , softmax represents the normalization operation, and \(\sum\) represents the summation operation. The final output \(X out is obtained by combining multiple propagation layers.

[0052] Apply Modified GCN to the \(k\)-truss subgraph to obtain features. Since the feature propagation and feature mapping in Modified GCN are carried out separately, it can effectively alleviate the overfitting problem. Therefore, Modified GCN has better performance in capturing high-order structural information.

[0053] Apply the Modified GCN to the subgraphs on each time slice, which is expressed by the formula:

[0054]

[0055] where k represents the subgraph with a granularity of k, and the value range of k is from 2 to K, where K is the subgraph with the largest granularity manually set by applying the k-truss algorithm, t represents the time slice index, and the range is from 1 to T, where T is the latest time slice. represents the node representation of the k-truss subgraph in the t-th time slice, represents the adjacency matrix of the k-truss subgraph in the t-th time slice. In this way, we obtain the node representations at all granularities and all time slices. On graph G t the multi-granularity sequence of node representations is At granularity k, the time sequence of node representations is

[0056] To better learn the structural features, assume that there is a potential structural evolution process in the multi-granularity subgraph sequence at each time. RNN is very suitable for processing highly correlated sequences. On graph G t reverse the sequence to and input it into the RNN to learn the structural information. Mathematically, it is expressed as:

[0057]

[0058]

[0059] …

[0060]

[0061] where represents the hidden state, represents the input feature. In the input sequence, the subgraph structure with a higher k value is denser and has fewer nodes. In this process, the structural information at different granularities is merged. The final hidden state output contains the structural attributes of the entire graph at time slice t. Perform this operation for each time slice and finally obtain the structural feature representation

[0062] Step 3. Multi-granularity dynamic feature learning.

[0063] In the time dimension, subgraphs with different granularities make different contributions to the evolution of the entire graph. Differentially processing subgraphs with different granularities helps to model the complete graph evolution process.

[0064] Use RNN to learn the dynamic evolution laws of subgraphs at different granularities. At each granularity k, is input into the RNN, which is formulated as follows:

[0065]

[0066]

[0067] …

[0068]

[0069] where T is the length of the time series, represents the hidden state. The dynamic feature learning module is similar to the structure fusion module, except for their inputs. At this time, the input sequence is a time series of different granularities, and the purpose is to capture the dynamic evolution features. Note that the structure embedding sequence is also input into the RNN and outputs the structure feature S T .

[0070] To obtain the final node representation, it is necessary to fuse the subgraph dynamic evolution features learned at multiple granularities. To obtain the importance of different granularities, a multi-head self-attention mechanism is introduced. The weights at each granularity are learned through the self-attention mechanism, and then the weights are multiplied by the node representations of the corresponding granularities and summed to obtain the fused node dynamic evolution feature representation.

[0071] The attention coefficient formula of node v i at the k granularity is as follows:

[0072]

[0073] where w k and w k’ are parameters, is the node representation, and the attention coefficient can be obtained after normalization. The representation of the fused node is:

[0074]

[0075] σ is the activation function, and the dynamic evolution feature representation H out of all the final nodes is obtained. The final node representation H emb includes the structure feature and the dynamic evolution feature, which is formulated as follows, where represents the matrix concatenation operation:

[0076]

[0077] Step 4. Predict future links.

[0078] Finally, update the network parameters according to the optimization function until the model converges and the performance on the test set reaches the optimal, and train the optimal model; perform the Hadamard product operation on the representations of the two nodes corresponding to the link, predict the link score through the node representations, and then train a logistic regression classifier to determine whether the link exists, so as to predict the graph G at the next time T T+1 on the links.

[0079] Next, taking the input dynamic graphs {G1, G2,..., G T} as an example, the present invention will be further described.

[0080] Figure 2 is the flowchart of dynamic link prediction, including input, multi-granularity graph partitioning, multi-granularity structural feature extraction, multi-granularity dynamic feature extraction and fusion, dynamic link prediction, and output.

[0081] Figure 3 is the schematic diagram of the dynamic link prediction architecture of the present invention. This architecture mainly consists of four parts: multi-granularity graph partitioning, multi-granularity structural feature extraction, multi-granularity dynamic feature extraction and fusion, and dynamic link prediction.

[0082] Step 1. Multi-granularity graph partitioning.

[0083] Apply the k-truss decomposition algorithm on each time slice of {G1, G2,..., G T} to obtain the corresponding multi-granularity graphs. As shown on the left, the multi-granularity subgraphs decomposed from Gt are Figure 3 where the superscript represents time and the subscript represents the serial number of the multi-granularity graph. According to the definition of k-truss, the subgraph sequence starts from the 2-truss subgraph (the original graph), and k_max is a manually set value. where the superscript represents time and the subscript represents the serial number of the multi-granularity graph. Since the definition of k-truss, the subgraph sequence starts from the 2-truss subgraph (the original graph), and k_max is a manually set value.

[0084] Step 2. Multi-granularity structural feature learning.

[0085] First, apply Modified GCN on the graph of each granularity at each time to extract features, and obtain Then, reverse the multi-granularity subgraph sequence at each time t and input it into the RNN to capture the potential structural features. Perform this operation on each time slice and finally obtain the structural feature representation

[0086] Step 3. Multi-granularity dynamic feature extraction and fusion

[0087] First, use the RNN to learn the dynamic evolution law of subgraphs of different granularities. At each granularity k, input it into the RNN to obtain the dynamic feature representation The dynamic features of multiple granularities are represented as Combined with the self-attention mechanism, calculate the importance degree (weight) of each granularity, multiply the dynamic feature representation by the weight and sum them to obtain the total dynamic feature representation. Finally, combine the structural features and the dynamic features to obtain the final node representation H emb .

[0088] Step 4. Dynamic link prediction.

[0089] Finally, update the network parameters according to the optimization function until the model converges and the effect on the test set reaches the optimal, and train the optimal model; perform the Hadamard product operation on the representations of the two nodes corresponding to the link, predict the link score through the representations of the nodes, and then train a logistic regression classifier to judge whether the link exists, so as to predict the graph G T+1 on the link at the next time T.

[0090] Experimental data: The experiment uses AUC as the evaluation index. AUC is an index to measure the learning ability of the model, which represents the probability of positive samples before negative samples, and the range is 0-1. The closer it is to 1, the better the model effect. The applied datasets UCI, AS, MATH, FACEBOOK, ASKU and ENRON are publicly available graph datasets, and the information is shown in Table 1.

[0091] Table 1. Dataset information

[0092] Dataset Number of nodes Number of edges Number of time slices UCI 1899 59835 7 AS 6828 1947704 100 MATH 24740 323357 77 FACEBOOK 60730 607487 27 ASKU 74924 356822 21 ENRON 87036 530284 38

[0093] Comparative experiments GCRN, DynGEM, DynAE, DynRNN, DynAERNN, EvloveGCN and CTGCN are all existing advanced dynamic link prediction methods, and the explanations are as follows:

[0094] GCRN: This model combines the GCN model and the RNN model to find the dynamic patterns in the graph and generate node embeddings in the dynamic graph.

[0095] DynGEM: This work is a graph embedding method based on deep autoencoding, which extends SDNE to dynamic graphs. This method can incrementally generate the embedding representation of the current graph snapshot based on the embedding representation of the previous graph snapshot, that is, use the embedding representation of the previous graph snapshot as the initial value, perform gradient training, and incrementally process the dynamic graph, and has very good performance.

[0096] Dyngraph2vec: This work mainly focuses on the features of dynamic graphs, proposes an embedding method to learn the evolving structure in dynamic graphs and predict invisible links. It uses multiple non-linear layers to learn the structural patterns in each network, and at the same time uses a recurrent layer to learn the temporal transitions in the network. This model has three variants DynAE, DynRNN and DynAERNN.

[0097] EvolveGCN: This framework is a dynamic variant of GCN. Instead of using node embedding, it uses GCN to learn the weights at a single moment in the time dimension. Meanwhile, considering the correlation of GCN weights between different moments, it uses RNN to evolve GCN parameters to capture the dynamics of the sequence.

[0098] CTGCN: This model proposes a temporal GCN, which combines GCN and RNN to capture dynamics while capturing the connection and structural information of the graph.

[0099] The experimental results are shown in Table 2. Some methods could not be experimented on large graphs due to memory limitations, indicated by "-".

[0100] Table 2. AUC Results of Dynamic Link Prediction Experiments

[0101]

[0102] Based on the same inventive concept, another embodiment of the present invention provides a dynamic link prediction system based on multi-granularity evolution using the above method, which includes:

[0103] A multi-granularity graph partitioning module for partitioning the graph under each time slice of the dynamic graph to obtain multi-granularity subgraphs;

[0104] A multi-granularity structural feature extraction module for extracting the structural features of nodes from the multi-granularity subgraphs;

[0105] A multi-granularity dynamic feature learning module for learning the dynamic evolution law of the graph from the multi-granularity subgraphs, obtaining the dynamic evolution features of subgraphs with different granularities, and fusing the dynamic evolution features of subgraphs with different granularities to obtain the dynamic evolution features of nodes;

[0106] A dynamic link prediction module for fusing the structural features of nodes with the dynamic evolution features of nodes to obtain node representations containing spatio-temporal features, and predicting future links based on the node representations containing spatio-temporal features.

[0107] Based on the same inventive concept, another embodiment of the present invention provides an electronic device (such as a computer, server, smart phone, etc.), which includes a memory and a processor. The memory stores a computer program, and the computer program is configured to be executed by the processor. The computer program includes instructions for performing each step in the method of the present invention.

[0108] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disc), and the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, each step of the method of the present invention is implemented.

[0109] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and implement it accordingly. Those of ordinary skill in the art can understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification, and the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A dynamic link prediction method based on multi-granularity evolution, characterized in that It includes the following steps: Partition the graph at each time slice of the dynamic graph to obtain multi-granularity subgraphs; Extract the structural features of nodes from the multi-granularity subgraphs; Learn the dynamic evolution law of the graph from the multi-granularity subgraphs, obtain the dynamic evolution features of subgraphs with different granularities, and fuse the dynamic evolution features of subgraphs with different granularities to obtain the dynamic evolution features of nodes; Fuse the structural features of nodes and the dynamic evolution features of nodes to obtain node representations containing spatio-temporal features, and predict future links based on the node representations containing spatio-temporal features; Use the k-truss subgraph decomposition algorithm to partition the graph at each time slice to obtain the multi-granularity subgraphs; Use a graph convolutional network based on disentangled propagation and mapping operations, and apply it to the k-truss subgraph to extract the structural features of the nodes; The graph convolutional network based on disentangled propagation and mapping operations has the following formula: att l = σ(MLP(X l )) l = 0, 1, 2, …, J Among them, l is the number of GCN layers; J is the maximum depth of propagation; is the node feature, N represents the number of nodes, d represents the feature dimension; MLP represents the fully connected operation; X l represents the node feature at depth l; is used to adaptively adjust the information that each node needs to retain at different propagation depths; represents att l Each element in it is multiplied by the corresponding d-dimensional node representation X J ; softmax represents the normalization operation; sum represents the summation operation; the final output X out is obtained by merging multiple propagation layers.

2. The method according to claim 1, wherein The learning of the dynamic evolution law of the graph from the multi-granularity subgraphs uses an RNN to learn the dynamic evolution law of subgraphs with different granularities.

3. The method according to claim 1, characterized in that, The fusion of the dynamic evolution features of subgraphs with different granularities to obtain the dynamic evolution features of nodes includes: introducing a multi-head self-attention mechanism, learning the weights at each granularity through the self-attention mechanism, and then multiplying the weights by the node representations corresponding to the granularities and summing them to obtain the fused node dynamic evolution feature representation.

4. The method according to claim 1, wherein The prediction of future links based on the node representations containing spatio-temporal features includes: performing a Hadamard product operation on the two node representations corresponding to the link, predicting the link score through the node representations, and then training a logistic regression classifier to determine whether the link exists, so as to predict the links on the graph at the next time.

5. A dynamic link prediction system based on multi-granularity evolution using the method described in any one of claims 1 to 4, characterized in that, It includes: A multi-granularity graph partitioning module for partitioning the graph at each time slice of the dynamic graph to obtain multi-granularity subgraphs; A multi-granularity structural feature extraction module for extracting the structural features of nodes from the multi-granularity subgraphs; A multi-granularity dynamic feature learning module for learning the dynamic evolution law of the graph from the multi-granularity subgraphs, obtaining the dynamic evolution features of subgraphs with different granularities, and fusing the dynamic evolution features of subgraphs with different granularities to obtain the dynamic evolution features of nodes; A dynamic link prediction module for fusing the structural features of nodes and the dynamic evolution features of nodes to obtain node representations containing spatio-temporal features, and predicting future links based on the node representations containing spatio-temporal features.

6. An electronic device, characterized in that, It includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the computer, the method according to any one of claims 1 to 4 is implemented.

Images