A method for analyzing and predicting electronic target sequences based on graph neural networks

Abstract

The present invention provides an electronic target sequence analysis and prediction method based on a graph neural network, which uses a graph convolutional neural network to analyze the electronic target sequence in a real-time scenario, wherein a graph convolutional layer is used to aggregate nodes and edge parameters to generate graph features, and a linear classifier is used to classify the electronic target sequence based on the graph features, thereby identifying the task label of the electronic target sequence; a graph autoencoder is used to analyze the electronic target sequence in a real-time scenario, wherein a graph convolutional neural network is used as an encoder to calculate the graph features, and negative links are randomly added, and finally the inner product is used to restore the original graph structure, thereby inferring the hidden target relationship of the electronic target sequence; a time-domain graph convolutional neural network is used to analyze the electronic target sequence with time series characteristics, wherein a graph convolutional neural network is used to calculate the graph features, and a gated recursive unit is used to calculate the electronic target sequence structure and parameters at the next moment based on the graph features, thereby predicting the possible future structure and parameter changes of the electronic target sequence.

Description

Technical Field

[0001] The present invention relates to analysis technology of graph data structures, and in particular to knowledge graph data analysis and prediction technology based on graph neural network. Background Art

[0002] Complex electromagnetic environments often contain multi-dimensional and multi-modal data such as multiple platform targets, electronic targets, coordination relationships, and electromagnetic parameters. Graphs, as a widely used data structure, are very suitable for describing such unstructured and internally interrelated data.

[0003] We call the above-mentioned data set with electromagnetic information and multi-dimensional complex information an electronic target sequence. The electronic target sequence describes in detail the combined information of platforms and equipment in a specific scene, and refers to the comprehensive representation of all targets and relationships in a scene. This sequence includes all platform targets, electronic targets and various parameters of targets in the scene, including various types of electronic equipment, such as radars, radio stations, etc., as well as various dimensional relationships between all nodes, such as command relationships, association relationships, communication relationships, and dynamic changes in sequences and other information.

[0004] After modeling and representing the electronic target sequence with a graph structure, it becomes a question to be analyzed and discussed to study and design methods to analyze and predict the electronic target sequence, effectively utilize the information of the electronic target sequence, and provide users with reference and warnings.

[0005] Electronic target sequence analysis focuses on the system operation of electronic target sequences. The main tasks include:

[0006] ·Association recognition: that is, judging which known electronic target sequence the electronic target sequence in the real-time scene belongs to, so as to know the label of the real-time electronic target sequence (such as task, subordinate, etc.);

[0007] Analysis and prediction: It is to determine whether there are undetected targets or composition relationships in the electronic target sequence of the real-time scene, whether there are abnormal targets or relationships, and predict the possible electronic targets and sequence structures.

[0008] Since knowledge graphs or simple directed graphs are generally used to model electronic target sequences, the algorithms used are generally the most prominent methods in graph data structure analysis: graph data matching.

[0009] Exact graph matching is the most widely used graph matching technology at present. It plays an important role in social network query, biological data analysis, social security analysis and other problems that require high accuracy of matching results. From the perspective of algorithm design, exact graph matching technology is divided into index-free matching technology and index-based matching technology. Index-free matching technology mainly adopts search strategy, which matches all nodes in the data graph in sequence by analyzing node attributes and node neighbor structure, and is suitable for accurate matching of small-scale data graphs. Ullmann algorithm is the earliest known index-free matching method. Representative algorithms include VF2, GraphQL, GADDI, Spath, etc. These algorithms mainly improve the search strategy by adding pruning, merging and other auxiliary information on the basis of Ullmann algorithm.

[0010] However, due to algorithmic limitations, such graph matching technology often only builds a database based on known information and data, and then associates and identifies electronic target sequences based on the database. Generally, an accurate graph matching algorithm is used to force the real-time electronic target sequence to match the known information, and then based on the known information, the possible node or edge relationship in real time is inferred. Its usage scenarios are very limited and are easily affected by the completeness of known information. Secondly, accurate graph matching algorithms such as the Ullmann algorithm focus more on the structure of the graph itself, ignoring the parameters of the nodes themselves, and cannot make good use of the multi-modal and multi-structured data in the scene.

[0011] At the same time, in actual scenarios, the structure of an electronic target sequence is often not static. Over time, the platform target coordinates, equipment target parameters, and communication and command relationships between targets may change, that is, electronic target sequences with timing characteristics. Current research is confined to the analysis and speculation of static electronic target sequences themselves, and lacks relevant research on such data analysis and prediction.

[0012] With the introduction of graph neural networks (GNN), many achievements have been made in machine learning analysis of non-Euclidean space data. Graph convolutional networks and graph attention networks have been widely used in chemistry, transportation and other fields. GNN can define convolution operations on graphs, aggregate multi-dimensional feature parameters of nodes, and use the relationship between nodes to perform graph classification, node classification and edge prediction on the overall graph structure. Therefore, it has a wide range of application prospects in non-Euclidean space data such as electronic target sequences with multi-dimensional features that can be represented by graph structures. Summary of the invention

[0013] The technical problems to be solved by the present invention are:

[0014] 1. In real-time scenarios, the association recognition task of electronic target sequences cannot fully utilize the parameters of the nodes themselves;

[0015] 2. When inferring hidden targets and relationships in electronic target sequences, it relies too much on known data and cannot function effectively when there is insufficient known data;

[0016] 3. When analyzing the electronic target sequence, there is a lack of analysis of the timing characteristics and a lack of prediction function for the timing electronic target sequence;

[0017] A method is proposed to model the electronic target sequence based on graph data structure, identify the electronic target sequence through association through GNN, infer the relationship between hidden target nodes and targets, and predict the structure of time-series electronic target sequence.

[0018] The technical solution adopted by the present invention to solve the above technical problems is a GNN-based electronic target sequence analysis and prediction method, comprising the following steps:

[0019] Modeling steps: Represent the data of the real-time scene to be analyzed with a graph data structure to form a real-time electronic target sequence. Use nodes to represent the real-time platform targets and electronic equipment targets. The attributes of the nodes are the parameters of the platform targets and electronic equipment targets. Use edges to represent the communication and command relationship between the platform targets and the electronic equipment targets. Forming the electronic target sequence completes the real-time data modeling. After that, encode the electronic target sequence to facilitate it as the input data of the neural network.

[0020] Association identification step: The electronic target sequence represented by the graph data structure is identified by the graph matching algorithm, and the known formation to which it is associated is used to complete the formation association; at the same time, the encoded electronic target sequence is input into the trained association identification network based on the graph neural network structure, which aggregates the edge and node parameter features of the real-time electronic target sequence through hierarchical calculation to perform task identification, and outputs the task label of the electronic target sequence to complete the sequence association;

[0021] Hidden node and relationship inference step: According to the task label and formation association results, the hidden target nodes are inferred based on the graph database; at the same time, the encoded electronic target sequence is input into the trained edge prediction network based on the graph autoencoder structure, and the network infers the hidden target relationship of the electronic target sequence through node embedding to obtain the relationship inference result; the edge prediction network based on the graph autoencoder structure uses the graph neural network as the encoder and the inner product as the decoder;

[0022] Time series structure prediction step: The encoded real-time electronic target sequence is input into the trained time series structure prediction network based on the time-graph convolutional network structure. The network predicts the electronic target sequence structure and parameters of the next beat according to the set time step and outputs them.

[0023] The present invention uses a graph convolutional neural network to analyze the electronic target sequence in a real-time scenario, wherein a graph convolutional layer is used to aggregate nodes and edge parameters to generate graph features, and a linear classifier is used to classify the electronic target sequence based on the graph features, thereby identifying the task label of the electronic target sequence. A graph autoencoder is used to analyze the electronic target sequence in a real-time scenario, wherein a graph convolutional neural network is used as an encoder to calculate the graph features, and negative links are randomly added, and finally the inner product is used to restore the original graph structure, thereby inferring the hidden relationship between targets in the electronic target sequence. A time-domain graph convolutional neural network is used to analyze the electronic target sequence with timing characteristics, wherein a graph convolutional neural network is used to calculate the graph features, and a gated recursive unit is used to calculate the electronic target sequence structure and parameters at the next moment based on the graph features, thereby predicting the possible future structural and parameter changes of the electronic target sequence.

[0024] The beneficial effects of the present invention are: in real-time scenarios, the association recognition task of the electronic target sequence can well and fully utilize the parameters of the node itself; it can infer the hidden targets and relationships of the electronic target sequence; and it can predict the structure and parameter changes of the electronic target sequence in combination with the timing characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flow chart of the implementation of the present invention;

[0026] Figure 2 An example of an electronic target sequence represented by a graph data structure;

[0027] Figure 3 A typical GCN network structure;

[0028] Figure 4 The neural network structure of the association recognition graph in the present invention;

[0029] Figure 5 The calculation process of the association identification network in the present invention;

[0030] Figure 6 A simple GAE structure diagram;

[0031] Figure 7 The network structure of the present invention uses GAE to perform edge prediction tasks;

[0032] Figure 8 Edge inference network calculation flow chart in the embodiment;

[0033] Fig. 9 A schematic diagram of the structure of T-GCN;

[0034] Fig.10 A flowchart of the network calculation for timing structure prediction in the embodiment. DETAILED DESCRIPTION

[0035] The flowchart of the present invention is as follows Figure 1 As shown, the specific steps are as follows:

[0036] S1: Modeling Step a) Real-time Data Modeling

[0037] The algorithm of the system of the present invention runs on a graph data structure, so the system first acquires real-time electronic target sequence data and performs graph data structure representation on the real-time electronic target sequence data.

[0038] In the embodiment, nodes are used to represent real-time platform targets and electronic equipment targets, and the attributes of the nodes are the parameters of the platform targets and the electronic equipment targets; edge relationships are used to represent the communication, command, association, subordination and other relationships between the platform targets and the electronic targets.

[0039] The parameters of the platform node, radar node and radio station node in the node make full use of the parameters that can be reported by the acquisition device, and then save them as the attributes of the target node. In the embodiment, the data parameter structure collected in the real-time scenario is required to be the same as the known parameter structure.

[0040] In the edge relationship, the command relationship refers to the relationship between two platform targets in the real-time scenario, in which one platform commands the other platform; the communication relationship refers to the relationship between two platform targets communicating with each other; the association relationship refers to some implicit coordination between platforms, but other relationships that do not belong to command or communication; the subordinate relationship refers to the mounting relationship between radar nodes and communication nodes and platform nodes. The command relationship and the association relationship can be obtained through the frequent item set algorithm, and the communication relationship and the subordinate relationship can be obtained through the detection device. In the embodiment, the above four relationships are defined as real-time known data.

[0041] For example, an electronic target sequence represented by a graph data structure such as Figure 2 As shown, all platform targets include parameters of “altitude”, “heading”, “speed”, “longitude and latitude”, “type”, “name” and “model”.

[0042] b) Real-time data encoding

[0043] After the real-time data is represented by a graph data structure, since the characteristics of some nodes and parameters are in string format, they need to be encoded before training and real-time recognition so that the data can be used as input for the neural network.

[0044] Since the string parameters of each individual are often independent of each other and there are not many types, the present invention uses a simple hash algorithm, the BKDRHash algorithm, to encode the string type parameters. The algorithm python code is as follows:

[0045]

[0046] For example, the string "AN / SPS" is encoded to obtain the value: 1009179012, which can be input into the subsequent neural network for training or real-time prediction.

[0047] In addition, the three parameters of platform type, node type and edge type are fixed and few in number, so they are coded in a fixed way, such as 0 for command relationship and 4 for radar node, etc., and written into the configuration items to ensure the consistency of the entire system. The task label of the entire sequence, as the label of the classification task of associated identification, belongs to the subsequent output content, and is encoded in order according to the number of tasks.

[0048] S2: Association Identification

[0049] a) Formation Identification

[0050] After modeling the real-time data and obtaining the real-time electronic target sequence represented by the graph data structure, formation identification is first performed, that is, the real-time electronic target sequence is associated with the formation in the known database to facilitate the subsequent analysis and calculation using the formation information.

[0051] When the names and model characteristics of the platform nodes, radar nodes and radio nodes in the real-time electronic target sequence are known, the direct node matching method is used to search the database to obtain all associated formations and output the formation number.

[0052] When there are unknown nodes, or when a node is detected but the node name / model is unknown, the precise graph matching algorithm, Ullmann algorithm, is used for formation identification.

[0053] The Ullmann algorithm is a classic algorithm for subgraph isomorphism problems. The following are the steps to use the Ullmann algorithm to determine whether graph Q is a subgraph of graph G:

[0054] 1. Label all nodes and edges in the query graph Q and the target graph G. The label of a node and an edge is a non-negative integer indicating their type. For example, if a node is a "roadbed" node, then it might be labeled 1.

[0055] 2. For each node v in the query graph Q, select an alternative set S(v), which contains all nodes marked as type v in the target graph G. If v has no alternative nodes, the algorithm can be stopped because the query graph Q has no subgraph that can be found isomorphic in the target graph G.

[0056] 3. For each edge e in the query graph Q, calculate the intersection of the candidate nodes of the two nodes it connects and obtain a candidate set C(e).

[0057] 4. For each node v in the query graph Q, recursively search each node w in the candidate set S(v) in the order of the nodes in the query graph Q. For each selected node w, if for all nodes adjacent to v in the query graph Q, there is an edge connecting v and the corresponding node of w in the target graph G, then add w to the candidate set S(v) and continue searching for the next node. If no candidate node can be found, it is necessary to backtrack and select other nodes.

[0058] 5. If all nodes in the query graph Q can be matched to nodes in the target graph G, it means that the query graph Q is a subgraph of the target graph G.

[0059] The time complexity of Ullmann's algorithm is O(n^|Q|), where n is the number of nodes in the target graph G and |Q| is the number of nodes in the query graph Q.

[0060] Finally, the number and information of the associated formation are also output.

[0061] b) Task Identification

[0062] After encoding the real-time electronic target sequence, it is input into the association recognition network, and the graph convolutional neural network GCN is used to perform the full-image classification task on the real-time electronic target sequence to obtain the task label of the real-time electronic target sequence.

[0063] The core idea of ​​GCN is to learn a function mapping f(.) on each layer of graph convolution, and map the nodes v in the graph through the mapping i can aggregate its own features x i With its neighbor feature x j ,j∈N(v i ) to generate node v i New representation.

[0064] A typical GCN network structure for classifying graph nodes is as follows: Figure 3 As shown. It contains an input layer, two hidden layers and an output layer, where each hidden layer uses a layer of graph convolution and a layer of ReLu activation layer to calculate the features of the graph. The principle of each GCN hidden layer is as follows:

[0065]

[0066] Among them, the input of the l-th layer network is H l , A is the adjacency matrix with self-connection added, D is the degree matrix, W l is the parameter to be trained, σ is the corresponding activation function, which is the final form of GCN.

[0067] In the embodiment, the main focus is on using GCN to perform supervised classification on graph data. The network model uses two graph convolution layers and one linear classifier layer. The model structure is as follows: Figure 4 shown.

[0068] After the encoded electronic target sequence diagram data structure enters the association recognition network, its operation process is as follows: Figure 5 As shown in the figure. The association recognition network includes two layers of graph neural network convolution layers and a linear classifier. The coded electronic target coding sequence first passes through two layers of graph neural network convolution layers. Each layer of graph convolution layers aggregates the features of neighboring nodes and adjacent edges to the current node, generates new graph data structure features of the current node, and finally inputs the graph data structure containing the latest parameters into a layer of linear classifier, outputs the corresponding probability of each task label, and selects the largest one as the final output task label.

[0069] S3: Hidden nodes and relationship inference

[0070] a) Hidden node inference

[0071] After formation identification and task identification in step 2, we can obtain the formation number and task label associated with the real-time electronic target sequence. We use the formation number as the primary index and the task label as the secondary index, and read the corresponding known database electronic target sequence associated with the graph database, so as to use the breadth-first search algorithm to infer the hidden target node.

[0072] Breadth-first search algorithm (also known as breadth-first search) is one of the simplest graph search algorithms. This algorithm is also the prototype of many important graph algorithms. Dijkstra's single-source shortest path algorithm and Prim's minimum spanning tree algorithm both use ideas similar to breadth-first search.

[0073] The present invention uses a layer of breadth-first search algorithm to output real-time platform targets or electronic equipment targets that may be hidden and not discovered by detection equipment in a real-time electronic target sequence.

[0074] b) Hidden edge relationship inference

[0075] The encoded real-time electronic target sequence is input into the trained edge prediction network, and the graph autoencoder GAE is used to infer the hidden target relationship of the real-time electronic target sequence, and the undetected target relationship that may exist in the real-time electronic target sequence is output.

[0076] like Figure 6As shown in the figure, the main purpose of GAE is to obtain suitable embeddings to represent the nodes in the graph so that they can be used in other tasks. GAE can reconstruct the embeddings of the nodes in the graph through the encoder-decoder structure to support the next task.

[0077] In the embodiment, GAE is used for edge prediction tasks, a graph convolutional neural network is used as an encoder, and an inner product is used as a decoder. The edge prediction network is as follows: Figure 7 As shown, it includes an encoder encoder containing two graph convolutional layers Graph Conv, a random negative link module Add negative links, a decoder decoder and a binary classifier.

[0078] The edge prediction network processing flow is as follows Figure 8 As shown in the figure, after inputting the encoded real-time electronic target sequence, the encoder in the model first creates a node embedding NodeEmbedding output graph embedding features through a graph Graph Conv with two convolutional layers; at the same time, negative links Add negative links are randomly added to the original graph, which makes the link prediction task become the positive links of the original edges and the negative links of the newly added edges to increment the electronic target sequence; the graph embedding features and the incremented electronic target sequence are input to the decoder decoder; the decoder decoder uses node embedding to perform link prediction (binary classification) on all edges (including negative links). It calculates the dot product Node pair multiplication of the node embedding from a pair of nodes on each edge, and then aggregates the value of the entire embedding dimension Aggregateembed dim to output the new graph data structure feature to the binary classifier. The binary classifier creates a value representing the probability of edge existence on each edge through the Sigmoid function and outputs the binary classification result. Through the binary classification results, we can output the possible command relationship, communication relationship and association relationship between platform targets in the real-time electronic target sequence.

[0079] S4: Temporal sequence structure prediction

[0080] The encoded real-time electronic target sequence is input into the trained temporal sequence structure prediction network, and the time-graph convolutional network T-GCN is used to predict the electronic target sequence structure and parameters of the next beat according to the set time step.

[0081] T-GCN is a temporal graph convolutional network that uniformly uses GCN and gated recurrent unit GRU in the framework. The former is used for graph node parameters and topological structure, and the latter is used to learn temporal features.

[0082] A T-GCN structure is as follows Fig. 9 It includes input layer, spatial layer, temporal layer and prediction layer, and consists of multiple time steps. In each time step, the network first passes the graph data through GCN convolution to calculate the feature vector of the graph, and then inputs the feature vector into the gated recurrent unit GRU. Based on the memory information of the past time step, it calculates and outputs the graph feature vector of the next time step and transmits the information of the current time step.

[0083] In the embodiment, the reporting period of the acquisition device is required to be less than or equal to 1 minute, and the parameters and topological structure of the real-time electronic target sequence change at the minute level or hour level. Therefore, the interval between each time-series electronic target sequence is set to 5 minutes, that is, the sampling period is 5 minutes, and the sequence timestamp interval is 300. At the same time, the network step size is set to 1, that is, the input of each time-series electronic target sequence predicts the structure of the electronic target sequence 5 minutes later. Finally, the number of hidden layer units in the network, that is, the hidden layer length is set to 100.

[0084] The processing flow of the time series structure prediction network is as follows Fig.10 As shown in the figure, after inputting the encoded real-time electronic target sequence, the trained graph convolution GCN unit is first used to calculate the graph structure data of the electronic target sequence to generate a new graph embedding feature embedding, and then the embedding is passed to the gated recurrent unit GRU unit. The GRU unit calculates the electronic target sequence features five minutes later based on the memory features of historical information and the current graph embedding feature embedding, and finally outputs the predicted electronic target sequence structure and parameters.

[0085] System Initialization

[0086] Before the system is run, it is necessary to initialize the system. Since it does not belong to the core content of the present invention, it is briefly described, which includes:

[0087] a) Initialization of the known formation database: Enter the formation information into the known information database to facilitate formation identification, formation structure and Figure 2 Since it is static information, the information it contains is less than some parameters of the electronic target sequence.

[0088] b) Initialization of sequence known information: Enter the sequence known information with task labels and time information into the database; associate the sequence with the formation to facilitate subsequent query as a cluster index;

[0089] c) Neural network training: The neural network involved in steps S2, S3, and S4 is trained based on the above-constructed data set, and a model parameter list is generated and saved.

[0090] The embodiment is based on GNN. First, the graph data structure of the real-time electronic target sequence is represented. Then, the real-time electronic target sequence is classified based on GCN, and its task label is obtained through association recognition. Node prediction is performed based on the label. At the same time, GAE is used to perform node embedding calculation on the real-time electronic target sequence to infer the possible hidden edge relationship between targets. Finally, T-GCN is used to predict the next sequence structure of the real-time electronic target sequence. This process has many advantages.

[0091] First, in the association recognition of electronic target sequences, the graph convolutional neural network is used to better and more fully utilize the multi-dimensional parameters of each target node in the scene, making the association recognition results more reliable and outputting the overall task labels of the electronic target sequences, which can facilitate users to refer to and judge.

[0092] Second, during association identification, formation identification is also performed, which is combined with the task label output in advantage one as a joint index, making the system more efficient in finding data and subsequent hidden target node inference tasks.

[0093] Third, in the inference of hidden edge relationships, the use of graph embedding networks effectively utilizes the topology and node parameter characteristics of the real-time electronic target sequence itself, so that relationship inference is not limited to known data, increases versatility, and provides users with more reference value.

[0094] Fourth, in the prediction of the structure of electronic target sequences, the use of time-domain graph convolutional neural networks can predict possible future structural changes of real-time electronic target sequences based on past data, rather than simply statically analyzing current targets and data, which can provide users with certain references and warnings.

Claims

1. A method for analyzing and predicting electronic target sequences based on graph neural networks, It is characterized in that The following steps are involved: Modeling steps: The data of the real-time scene to be analyzed is represented by a graph data structure to form a real-time electronic target sequence. Nodes are used to represent the real-time platform targets and electronic equipment targets. The attributes of the nodes are the parameters of the platform targets and electronic equipment targets. Edges are used to represent the communication and command relationship between the platform targets and the electronic equipment targets. The real-time data modeling is completed by forming an electronic target sequence. After that, the electronic target sequence is encoded to facilitate the use as input data for the neural network. Association identification step: The electronic target sequence represented by the graph data structure is identified by the graph matching algorithm, and the known formation to which it is associated is used to complete the formation association; at the same time, the encoded electronic target sequence is input into the trained association identification network based on the graph neural network structure, which aggregates the edge and node parameter features of the real-time electronic target sequence through hierarchical calculation to perform task identification, and outputs the task label of the electronic target sequence to complete the sequence association; Hidden node and relationship inference step: According to the task label and formation association results, the hidden target nodes are inferred based on the graph database; at the same time, the encoded electronic target sequence is input into the trained edge prediction network based on the graph autoencoder structure, and the network infers the hidden target relationship of the electronic target sequence through node embedding to obtain the relationship inference result; the edge prediction network based on the graph autoencoder structure uses the graph neural network as the encoder and the inner product as the decoder; Time series structure prediction step: The encoded real-time electronic target sequence is input into the trained time series structure prediction network based on the time-graph convolutional network structure. The network predicts the electronic target sequence structure and parameters of the next beat according to the set time step and outputs them.

2. The method according to claim 1, It is characterized in that The association recognition network based on the graph neural network structure includes two graph neural network convolutional layers and a linear classifier. The processing process of the association recognition network is as follows: After the encoded electronic target sequence enters the association recognition network, it first passes through two layers of graph neural network convolution layers. Each layer of the graph convolution layer aggregates the features of neighboring nodes and adjacent edges to the current node, generating a new graph data structure feature for the current node. Finally, the graph data structure containing the latest parameters is input into the linear classifier. The linear classifier outputs the corresponding probabilities of each task label, and selects the largest one as the final output task label.

3. The method according to claim 1, It is characterized in that The edge prediction network based on the graph autoencoder structure includes an encoder with two graph convolutional layers, a negative link addition module, a decoder, and a binary classifier; the processing process of the edge prediction network is as follows: The encoder creates node embedding and outputs graph embedding features for the encoded electronic target sequence through two graph convolutional layers; at the same time, the negative link adding module randomly adds negative links to the electronic target sequence of the original graph to increment the electronic target sequence; the graph embedding features and the incremented electronic target sequence are input into the decoder; the decoder calculates the dot product of the graph embedding features of a pair of nodes on each edge of the input, and then aggregates the values ​​of the entire graph embedding feature dimension, and outputs the new graph data structure features to the binary classifier. The binary classifier creates a value representing the probability of the existence of the edge on each edge through the Sigmoid function and outputs the binary classification result.

4. The method according to claim 1, It is characterized in that The temporal structure prediction network includes graph convolution units and gated recurrent units; the processing process of the temporal structure prediction network is: The graph convolution GCN unit calculates the graph structure data of the electronic target sequence according to the input encoded electronic target sequence to generate graph embedding features, and then passes the graph embedding features to the gated recursive unit. The gated recursive unit is based on the memory features of historical information and combined with the current graph embedding features. Calculate the characteristics of the electronic target sequence after a preset time, and finally output the predicted electronic target sequence structure and parameters.

Images