Network link prediction method and device, electronic equipment and medium

Abstract

The invention discloses a network link prediction method and device, electronic equipment and a medium. The method comprises the steps that a generation network G generates a state vector of a current time step according to a topological structure and node attributes of each independent snapshot of a dynamic network; the generative network G generates a predictive map snapshot # imgabs0 # of the next time step based on the state vector of the current time step, so that the discriminating network D performs discriminating training based on the predictive map snapshot # imgabs1 # and the real snapshot A [tau] + 1 until the predictive map snapshot generated by the generative network G passes the discriminating of the discriminating network D; according to the network link prediction method disclosed by the invention, the prediction map snapshot of the next time step can be predicted through the generation network G on the basis of the topological characteristics and node attributes of the current network, and the generation network G and the judgment network D are trained on the basis of the prediction map snapshot and the real snapshot; and the trained generative network G generates a prediction map snapshot with high precision to realize link prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication networks, and specifically relates to a network link prediction method, device, electronic equipment and medium. Background Art

[0002] In UAV networks, link topology prediction is a critical task, and the application of machine learning algorithms in complex networks has made significant progress. However, there are certain limitations in directly applying machine learning algorithms to network analysis.

[0003] Traditional link prediction methods are mainly divided into the following two categories: Node embedding-based methods: This type of method uses similarity indicators between nodes (such as the number of common neighbors, Jaccard similarity, etc.) to predict the probability of connection between nodes. For example, the CN algorithm and the Katz algorithm both use the similarity indicator of the number of common neighbors to predict links. This type of method is simple and easy to implement, but it cannot effectively capture the pattern of dynamic changes in the network, resulting in low prediction accuracy. Node embedding-based methods: This type of method maps nodes to a low-dimensional vector space, and then uses the distance or similarity between node vectors to predict the probability of connection between nodes. For example, the Node2vec algorithm, a node embedding method based on random walks, learns the representation of nodes through random walks. This type of method can effectively capture the structural characteristics of the network, but cannot effectively capture the pattern of dynamic changes in the network, resulting in limitations in prediction accuracy.

[0004] Therefore, in response to the above technical problems, it is necessary to provide a network link prediction method, device, electronic device and medium. Summary of the invention

[0005] The purpose of the present invention is to provide a network link prediction method, device, electronic device and medium, which can solve the technical problem that traditional link prediction methods cannot effectively capture dynamic changes in the network, resulting in low prediction accuracy.

[0006] In order to achieve the above object, a specific embodiment of the present invention provides a network link prediction method, the network link prediction method comprising:

[0007] The generated network G generates the state vector of the current time step according to the topological structure and node attributes of each individual snapshot of the dynamic network;

[0008] The generative network G generates a snapshot of the prediction graph for the next time step based on the state vector of the current time step For the discriminant network D to be based on the predicted graph snapshot and the real snapshot A τ+1 Perform identification training until the prediction image snapshot generated by the generator network G passes the identification of the discriminant network D;

[0009] The generative network G updates the predicted graph snapshot after being trained by discrimination

[0010] In one or more embodiments of the present invention, the topological structure and node attributes of the individual snapshots are represented by an adjacency matrix and the feature matrix Indicates that the adjacency matrix is set to generate the graph snapshot input of the network G, the adjacency matrix is set to a preset range, the feature matrix is set to generate the noise of network G.

[0011] In one or more embodiments of the present invention, the generation network G includes a GAE layer, an LSTM layer and a fully connected output layer.

[0012] The method further comprises:

[0013] The GAE layer is based on a sequence of graph snapshots and noise Z, the output represents the sequence

[0014] The LSTM layer outputs a sequence based on Output state vector And the last state vector h τ Output to the fully connected output layer;

[0015] The fully connected output layer is based on the hidden state h τ Generate a snapshot of the prediction graph for the next time step

[0016] In one or more embodiments of the present invention, the method further includes: the LSTM unit takes the input vector x of the current time step t and the state vector h of the previous time step t-1 As input, it then outputs the state vector h of the current time step t , with the final hidden state h τ+1 is considered as a distribution representation of historical snapshots and fed into the fully connected output layer.

[0017] In one or more embodiments of the present invention, the method further comprises:

[0018] The discriminant network D and the generative network G are trained. The training purpose of the discriminant network D is to distinguish the predicted image snapshots. and the real snapshot A τ+1 The training purpose of the generative network G is to predict the graph snapshot Through the identification of the discriminant network D, until the discriminant network D cannot distinguish the predicted image snapshot and the real snapshot A τ+1 .

[0019] In one or more embodiments of the present invention, the method further comprises:

[0020] The discriminant network D alternately inputs the prediction graph snapshots Or a real snapshot A τ+1 , and reshape it into the corresponding row vector;

[0021] The discriminant process of the discriminant network D includes:

[0022]

[0023] Where A′ represents the corresponding prediction graph snapshot reshaped into a row vector and the real snapshot A τ+1 ; as well as are the parameters of the hidden layer and output layer of the discriminant network D respectively; σ(·) represents the Sigmoid activation function of the hidden layer.

[0024] In one or more embodiments of the present invention, the prediction graph snapshot and the real snapshot A τ+1 The row-by-row length vector correspondence is that both are standardized to be within a preset range.

[0025] In another aspect of the present invention, a network link prediction device is provided, the device comprising:

[0026] Generative model, used to generate the state vector of the current time step of the network G according to the topological structure and node attributes of each individual snapshot of the dynamic network;

[0027] The training model is used for the generative network G to generate a snapshot of the prediction graph for the next time step based on the state vector of the current time step. For the discriminant network D to be based on the predicted graph snapshot and the real snapshot A τ+1 Perform identification training until the prediction image snapshot generated by the generator network G passes the identification of the discriminant network D;

[0028] Update the model for the generative network G to update the predicted graph snapshot after being trained by identification

[0029] In another aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory, wherein the memory stores instructions, and when the instructions are executed by the at least one processor, the at least one processor executes a network link prediction method.

[0030] In another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the network link prediction method are implemented.

[0031] Compared with the prior art, the network link prediction method of the present invention can predict the prediction graph snapshot of the next time step based on the topological characteristics and node attributes of the current network by generating a network G, and train the generated network G and the judgment network D based on the prediction graph snapshot and the real snapshot. After the training is completed, the generated network G generates a prediction graph snapshot with high precision to realize link prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 This is a schematic diagram of a network link structure according to an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of a network link prediction model according to an embodiment of the present invention

[0035] Figure 3 A flowchart of a network link prediction method according to an embodiment of the present invention;

[0036] Figure 4 A schematic diagram of an LSTM layer and a fully connected output layer according to an embodiment of the present invention;

[0037] Figure 5 A statistical graph of the AUC index of the experimental results of the network link prediction method of the present invention;

[0038] Figure 6 It is a statistical graph of the MAP index of the experimental results of the network link prediction method of the present invention;

[0039] Figure 7 It is a statistical chart of the ErrorRate indicator of the experimental results of the network link prediction method of the present invention;

[0040] Figure 8A module diagram of a network link prediction device according to another embodiment of the present invention;

[0041] Fig. 9 This is a hardware structure diagram of a computing device according to another embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0043] like Figure 1 As shown in the figure, the application scenario of the present invention is introduced. The cloud-edge-end network resource elastic scheduling architecture based on drones adopts the Fat-Tree topology structure; drones are integrated into the cloud layer, edge layer and end layer to build a comprehensive and layered computing and communication framework. In the cloud layer application, drones are used for data processing and storage to replace the functions of traditional data centers; drones in this layer mainly focus on performing large-scale data analysis tasks, using their rapid deployment capabilities to provide instant computing resources for specific areas or tasks. The edge layer drones are closer to ground users or sensor networks, which are used for data preprocessing and real-time analysis. At this level, drones adjust their positions and assigned tasks according to actual network conditions and data processing requirements through flexible scheduling to ensure low latency and high efficiency of data processing. Edge; At the end layer, the deployment of drones is further decentralized and specialized. End-layer drones interact with ground devices or users, are responsible for collecting data and performing simple edge computing tasks. The deployment of end-layer drones greatly improves the network coverage and the real-time nature of data collection. This architecture uses the hierarchical characteristics of the Fat-Tree topology to effectively allocate and optimize network resources by deploying drone nodes at different levels. This structure can adapt to the dynamic changes of the drone network and achieve flexible resource scheduling while ensuring network performance.

[0044] like Figure 2 As shown, a specific implementation of a network link prediction method based on the cloud-edge network structure of the above-mentioned drone is introduced, and the method specifically includes the following steps:

[0045] S101, generating a network G to generate a state vector of the current time step according to the topological structure and node attributes of each individual snapshot of the dynamic network.

[0046] In this embodiment, a link prediction model is used to implement link prediction of a cloud-edge network based on a drone. The link prediction model includes a generation network G and a judgment network D, wherein the generation network G includes a GAE (graph autoencoder) layer, an LSTM (long short-term memory network) layer, and a fully connected output layer; in this embodiment, the GAE layer uses an encoder to model the local topological structure of each individual graph snapshot in the dynamic network, such as Figure 3 As shown, the GAE layer includes an encoder and a decoder. In this embodiment, the encoder is an efficient variant of a convolutional neural network that can operate directly on the graph.

[0047] In one embodiment, the encoder represents the topological structure and node attributes of the graph snapshot in a low-dimensional manner, specifically including: and the feature matrix Represents the topological structure and node attributes of a single snapshot, adjacency matrix is set to generate the graph snapshot input of the network G, the adjacency matrix is set to a preset range, the feature matrix (where the i-th row of Z corresponds to the feature vector of node i) is set to generate the noise of the network G, where N is represented as N nodes in the static graph, each node has M-dimensional features or attributes.

[0048] Specifically, the encoder takes the feature matrix Z as input and performs a spectral graph convolution operation on it (local first-order approximation based on the adjacency matrix A), and the final output is generated in the same way as the standard fully connected layer. The overall operation of the specific encoder can be briefly defined as follows:

[0049]

[0050] in, is an approximate graph convolution filter, where (I N is an N-dimensional identity matrix), and W represents the weight matrix; f(·) is the activation function; and X represents the output of the encoder.

[0051] like Figure 2 As shown, for the temporal connection prediction task of multiple graph snapshots, the link prediction model inputs A for each graph snapshot t (t∈{τ-l,…,τ}) maintains an encoder GCN(Z,A t )=X t , the feature matrix is set to generate the noise input of the network G, where the value of Z is generated according to a certain probability distribution p, which can be a uniform distribution in this embodiment; in this embodiment, based on the feature matrix Z and the graph snapshot input The output of the GAE layer can be expressed as The GAE layer outputs the output of the encoder to the LSTM layer.

[0052] In one embodiment, if Figure 3 As shown in Figure 2, the LSTM layer has a strong ability to learn long-term dependencies of sequence data to capture the evolution pattern of weighted dynamic networks. The standard LSTM architecture can be described as a package unit with multiple multiplication gate units; the sequence representation based on the output of the GAE layer For the current time step t, the LSTM layer receives the input vector x of the current time step t and the state vector h of the previous time step t-1 , output the state vector h of the current time step t , where the state vector h at the current time step is t satisfy:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] h t =o t ☉tanh(s t );

[0059] Among them, i t 、f t , o t and t They represent input gate, forget gate, output gate and memory cell respectively;

[0060] {W x ,W h ,b} are the parameters of the corresponding unit; σ(·) is the sigmoid activation function; ⊙ represents element-by-element multiplication. Finally, the final hidden state h τ+1 is considered as a distribution representation of historical snapshots and fed into the fully connected output layer.

[0061] In a further embodiment, since the LSTM framework has the ability to learn the temporal information of sequence data, the LSTM framework (with multiple inputs and a single output) can be directly used to process the time link prediction task. However, there are still some limitations for the prediction of weighted dynamic networks, so the LSTM is trained using the mean square error (MSE) loss function to learn the temporal information of the dynamic network.

[0062] S102, the generation network G generates a snapshot of the prediction graph for the next time step based on the state vector of the current time step For the discriminant network D to be based on the predicted graph snapshot and the real snapshot A τ+1 Do identification training until the prediction image snapshot generated by the generative network G passes the identification of the discriminative network D.

[0063] In one embodiment, the fully connected output layer in the generative network G is based on the state vector h of the current time step t To generate a snapshot of the prediction graph for the next time step (represented in the form of corresponding row-by-row vectors), the fully connected output layer predicts the graph snapshot Output to the discriminant network D, based on the predicted graph snapshot and the real snapshot A τ+1 Train the discriminant network D and the generative network G. The purpose of training the discriminant network D is to distinguish the predicted image snapshots. and the real snapshot A τ+1 The training purpose of the generative network G is to predict the graph snapshot Through the identification of the discriminant network D, until the discriminant network D cannot distinguish the predicted image snapshot and the real snapshot A τ+1 .

[0064] In one embodiment, the training process of the generative network G and the adversarial network D can be expressed as:

[0065]

[0066] Where x is the input data from the training set, and z represents the noise generated by a certain probability distribution p(z) (uniform distribution). The present invention trains the generative network G through the adversarial process between the generative network G and the discriminative network D; the training purpose of the discriminative network D is to distinguish the snapshots of the prediction graph. and the real snapshot A τ+1 The training purpose of the generative network G is to predict the graph snapshot Through the identification of the discriminant network D, until the discriminant network D cannot distinguish the predicted image snapshot and the real snapshot A τ+1 .

[0067] In this embodiment, each input data of the fully connected output layer is usually represented as a vector (rather than a matrix form). When it is fed to the discriminant network D, based on the matrix input (i.e., the prediction graph snapshot Or a real snapshot A τ+1) is reshaped into corresponding row-by-row long vectors, and the WassersteinGAN (WGAN) framework is used to train the model. The output layer is set as a linear layer to directly generate output without nonlinear activation functions. The details of the discriminant network D can be expressed as follows:

[0068]

[0069] Where A′ represents the corresponding reshaped row vector or A τ+1 ; as well as are the parameters of the hidden layer and the output layer respectively; σ(·) represents the Sigmoid activation function of the hidden layer.

[0070] Given a real snapshot A τ+1 The edge weights of may have a large range of values ​​(e.g., [0, 2, 000]). τ+1 As the input of the discriminative network D, we take the real snapshot A τ+1 The element values ​​of are normalized to the range [0, 1]. The predicted graph snapshot given by the generative network G It is defined in the range [0, 1] and can therefore be used directly as input to the discriminative network.

[0071] S103, the generated network G updates the predicted graph snapshot after passing the identification training

[0072] In this embodiment, the generated network G includes a GAE layer, an LSTM layer, and a fully connected output layer. The GAE layer is based on the graph snapshot sequence. and noise Z, the output represents the sequence The LSTM layer represents the sequence based on the output Output state vector And the last state vector h τ Output to the fully connected output layer; the fully connected output layer is based on the hidden state h τ Generate a snapshot of the prediction graph for the next time step

[0073] The network link prediction method provided by the present invention is experimentally tested as follows:

[0074] The experimental test platform of the present invention adopts the AirSim simulation platform to simulate complex network topology and UAV flight environment, and uses the classic link prediction evaluation indicators AUC, MAP (mean average precision) and ErrorRate to verify the model of the present invention.

[0075] To verify the effectiveness of the link prediction method of the present invention. The network link prediction method of the present invention is compared with the classic CN and Katz algorithms, the node embedding-based method Node2vec, and the graph embedding-based method GAE-LSTM to verify the effectiveness of the link prediction method of the present invention. Among them, the link prediction method based on the similarity index uses the node similarity as the probability of generating a connection between nodes. To ensure the rationality of the experiment, all comparative experiments will be conducted under the data sample of the optimal model parameters of the present invention.

[0076] Among them, AUC satisfies:

[0077]

[0078] Among them, M is the number of comparisons, M' is the number of times the score of the selected edge in the positive example is greater than the score of the selected edge in the negative example, and M" is the number of times the two are equal. The larger the M' and M", the greater the probability that the positive example score is higher than the negative example score, and the higher the prediction accuracy.

[0079] Further MAP satisfies:

[0080]

[0081] Among them, AP is the area under the PR (precision-recall) curve.

[0082]

[0083] Among them, the set of nodes v and the connections between them are represented as {q.=1} MAP, which is used to measure the accuracy of the prediction results.

[0084] Error Rate is the ratio of the wrong connections LNmls to all the real connections LNe. Different from the SumD metric that only calculates the absolutely different connections in the two graphs, Error Rate considers the number of real connections. ErrorRate satisfies:

[0085]

[0086] To ensure the rationality of the experiment, all comparative experiments will be conducted under the data samples of the optimal model parameters of the present invention. Each group of experiments is sampled 20 times, and the corresponding AUC, MAP and Error Rate are calculated. Among them, the static network link prediction methods CN, Katz, and Node2vec predict the network link status at the next moment by analyzing the network status in the current snapshot.

[0087] like Figure 5As shown in the figure, the prediction effects of different link prediction models under the same environment reflected by AUC. It can be observed that the link prediction methods based on deep learning, Node2vec-LSTM, GAE-LSTM and GAE-GAN of the present invention are more stable than other methods. Overall, the AUC value of the network link prediction method under 20 test samples is always higher than that of other methods, and the prediction effects of CN and Katz are the worst.

[0088] like Figure 6 As shown in the figure, the MAP obtained by different link prediction methods shows that except for the three deep learning-based link prediction methods, the MAP of other methods is below 0.5. Even though Node2vec-LSTM reached 0.5 in the first case, the generalization performance of the model is low. The effects of different prediction methods reflected by the Error Rate indicator also show that except for the three deep learning-based prediction methods, no good prediction results were achieved. Among them, Node2vec-LSTM has a more stable effect than Node2vec-LRCV, reflecting that deep learning models have better generalization than traditional machine learning models.

[0089] like Figure 7 As shown in the figure, the network link prediction method of the present invention has good performance in AUC, MAP and Error Rate, and it can be observed from the fluctuation of the curve in the figure that the method of the present invention is more stable than other methods in the 20 sampling results. Compared with the GAE-LSTM model, the network link prediction method is improved by 0.87%, MAP is improved by 3.71%, and Error Rate is reduced by 0.45. It can be seen that the model can better learn the topological evolution law of the drone self-organizing network.

[0090] like Figure 8 As shown, a device for implementing a network link prediction method provided in a specific embodiment of the present invention is introduced.

[0091] In an embodiment of the present invention, a network link prediction device is provided, comprising:

[0092] Generative model 201, used to generate a state vector of the current time step of the network G according to the topological structure and node attributes of each individual snapshot of the dynamic network;

[0093] Training model 202, used to generate a prediction graph snapshot of the next time step generated by the network G based on the state vector of the current time step For the discriminant network D to be based on the predicted graph snapshot and the real snapshot A τ+1 Perform identification training until the prediction image snapshot generated by the generative network G passes the identification of the discriminative network D;

[0094] Update model 203, used to generate a snapshot of the prediction graph of the network G after the discriminative training

[0095] Fig. 9 According to an embodiment of this specification, a hardware structure diagram of a computing device 30 for network link prediction is shown. Fig. 9 As shown, the computing device 30 may include at least one processor 301, a memory 302 (e.g., a non-volatile memory), a memory 303, and a communication interface 304, and the at least one processor 301, the memory 302, the memory 303, and the communication interface 304 are connected together via a bus 305. At least one processor 301 executes at least one computer-readable instruction stored or encoded in the memory 302.

[0096] It should be understood that the computer executable instructions stored in the memory 302, when executed, cause at least one processor 301 to perform the above combined operations in various embodiments of the present specification. Figure 1-8 Describes the various operations and functions.

[0097] In the embodiments of the present specification, the computing device 30 may include, but is not limited to, a personal computer, a server computer, a workstation, a desktop computer, a laptop computer, a notebook computer, a mobile computing device, a smart phone, a tablet computer, a cellular phone, a personal digital assistant (PDA), a handheld device, a messaging device, a wearable computing device, a consumer electronic device, and the like.

[0098] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0100] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0102] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0103] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A network link prediction method, It is characterized in that The method comprises: The generated network G generates the state vector of the current time step according to the topological structure and node attributes of each individual snapshot of the dynamic network; The generative network G generates a snapshot of the prediction graph for the next time step based on the state vector of the current time step For the discriminant network D to be based on the predicted graph snapshot and the real snapshot A τ+1 Perform identification training until the prediction image snapshot generated by the generator network G passes the identification of the discriminant network D; The generative network G updates the predicted graph snapshot after being trained by discrimination 2. The network link prediction method according to claim 1, It is characterized in that The topological structure and node attributes of each snapshot are represented by the adjacency matrix and the feature matrix Indicates that the adjacency matrix is set to generate the graph snapshot input of the network G, the adjacency matrix is set to a preset range, the feature matrix is set to generate the noise of network G.

3. The network link prediction method according to claim 2, It is characterized in that The generating network G includes a GAE layer, an LSTM layer and a fully connected output layer. The method further comprises: The GAE layer is based on a sequence of graph snapshots and noise Z, the output represents the sequence The LSTM layer outputs a sequence based on Output state vector And the last state vector h τ Output to the fully connected output layer; The fully connected output layer is based on the hidden state h τ Generate a snapshot of the prediction graph for the next time step 4. The network link prediction method according to claim 3, It is characterized in that The method further includes: the LSTM unit takes the input vector x of the current time step t and the state vector h of the previous time step t-1 As input, it then outputs the state vector h of the current time step t , with the final hidden state h τ+1 is considered as a distribution representation of historical snapshots and fed into the fully connected output layer.

5. The network link prediction method according to claim 1, It is characterized in that The method further comprises: The discriminant network D and the generative network G are trained. The training purpose of the discriminant network D is to distinguish the predicted image snapshots. and the real snapshot A τ+1 The training purpose of the generative network G is to predict the graph snapshot Through the identification of the discriminant network D, until the discriminant network D cannot distinguish the predicted image snapshot and the real snapshot A τ+1 .

6. The network link prediction method according to claim 5, It is characterized in that The method further comprises: The discriminant network D alternately inputs the prediction graph snapshots Or a real snapshot A τ+1 , and reshape it into the corresponding row vector; The discriminant process of the discriminant network D includes: Where A′ represents the corresponding prediction graph snapshot reshaped into a row vector and the real snapshot A τ+1 ; as well as are the parameters of the hidden layer and output layer of the discriminant network D respectively; σ(·) represents the Sigmoid activation function of the hidden layer.

7. The network link prediction method according to claim 6, It is characterized in that The forecast graph snapshot and the real snapshot A τ+1 The row-by-row length vector correspondence is that both are standardized to be within a preset range.

8. A network link prediction device, It is characterized in that The device comprises: Generative model, used to generate the state vector of the current time step of the network G according to the topological structure and node attributes of each individual snapshot of the dynamic network; The training model is used for the generative network G to generate a snapshot of the prediction graph for the next time step based on the state vector of the current time step. For the discriminant network D to be based on the predicted graph snapshot and the real snapshot A τ+1 Perform identification training until the prediction image snapshot generated by the generator network G passes the identification of the discriminant network D; Update the model for the generative network G to update the predicted graph snapshot after being trained by identification 9. An electronic device, It is characterized in that include: at least one processor; and a memory storing instructions, which, when executed by the at least one processor, causes the at least one processor to execute the network link prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the network link prediction method according to any one of claims 1 to 7 are implemented.

Images