An opportunistic network link prediction method based on a time series graph neural network

Abstract

The application discloses a kind of opportunity network link prediction method based on timing diagram neural network, specifically related to timing diagram neural network technical field, including the following steps: S1, data acquisition and network snapshot generation;S2, weighted interface feature coding aggregation;S3, GAT and LSTM are combined to process embedding;S4, multilayer perception algorithm probability;S5, according to link construction loss function;S6, set training adjusts parameter;S7, multi-index evaluation optimization model.The application accurately captures the space-time characteristics of nodes, embedding, coding, decoding layer collaborative operation, improves the accuracy and stability of link prediction, stabilizes the reliability of network communication and the success rate of data transmission, innovates loss function and optimization strategy, solves the problem of data sparsity and overfitting, enhances the generalization ability, the model adapts to multi-domain network, accurately predicts the link in intelligent transportation, disaster rescue and other scenarios, optimizes resource allocation and communication efficiency, and effectively promotes the innovation and development of multi-industry network application.

Description

Technical Field

[0001] The present invention relates to the technical field of time sequence graph neural network, and in particular to an opportunistic network link prediction method based on time sequence graph neural network. Background Art

[0002] With the rapid development of wireless communication technology and equipment, traditional base station-dependent wireless communication networks, while widely used, are vulnerable. This led to the emergence of MANETs, ​​which enable self-organizing communication. However, early routing protocols were limited in dynamic scenarios. To address this dilemma, DTNs and their derivatives, opportunistic networks, emerged. The latter, with their unique advantages, have found widespread application in various fields, with their prominence in the military. Link prediction is crucial for opportunistic networks, but it faces numerous challenges. These include network dynamics rendering traditional methods ineffective, data sparsity hindering prediction, feature extraction hindered by complex factors, the need to balance efficiency and universality in predictive model design, and data fusion limited by format and information conflicts. These challenges urgently need to be overcome.

[0003] Chinese patent publication CN106934489B discloses a method for predicting time-series links in complex networks. This method leverages network evolution information to achieve highly accurate link predictions, with the core steps designed based on a holistic synchronous parallel computing model. The method is highly universal and applicable to predicting time-series links in a variety of social networks. Furthermore, it exhibits excellent scalability and is applicable to predicting time-series links in distributed environments.

[0004] However, the above methods are shallow in feature mining and only rely on interaction time and frequency. The algorithm lacks flexibility and is difficult to adapt to diverse networks. Summary of the Invention

[0005] The main purpose of the present invention is to provide an opportunistic network link prediction method based on a temporal graph neural network, which can effectively solve the problems of shallow feature mining, relying only on interaction time and frequency, and lack of algorithm flexibility.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for predicting opportunistic network links based on a temporal graph neural network comprises the following steps:

[0008] S1: Comprehensively collect multi-channel data within the opportunistic network, organize the data into a series of network snapshots according to predetermined time span rules, construct a weighted adjacency matrix to reflect the details of node connection interactions, and after normalization, lay a solid foundation for the underlying data architecture for link prediction.

[0009] S2: After accepting the weighted adjacency matrix and node feature materials, the encoding fusion operation is used to enhance the richness of feature dimensions. Attention calculation is used to accurately focus on the key features of the nodes. Weighted aggregation is used to generate an embedding matrix to capture the structural characteristics of the nodes and transfer the encoding process to deeply extract the feature connotations.

[0010] S3: Starting with node embedding materials, the spatial structure between nodes is mined with the help of graph attention networks, and the dynamic evolution trajectory of the node historical feature sequence is analyzed using long-short-term memory networks. The hidden states are spliced ​​and condensed into spatiotemporal embedding representations. The self-attention mechanism is used to dynamically adjust the weights according to the node coreness and time context, and the weights are transmitted to the decoding link after optimization.

[0011] S4: Start the multi-layer perceptron architecture, introduce the spatiotemporal embedding results of the coding layer into the input layer, use the deep abstract conversion function of the hidden layer, rely on the sophisticated processing of the activation function, and through multi-layer progressive conversion, map the features into a link generation probability matrix, clarify the probability of node emergence of the link, and provide a core decision-making reference for link prediction.

[0012] S5: Build a core architecture of the loss function around the predicted and real links, formulate a dynamic weighting strategy to address the link sparsity dilemma, integrate complexity management technology and label optimization methods, and use the cross-validation mechanism to fine-tune the weight parameters to achieve an ideal balance between model prediction accuracy and generalization breadth, thereby consolidating the model performance foundation.

[0013] S6: The dataset is divided into training, validation, and testing modules in appropriate proportions. Simulated annealing (to expand search breadth) and stochastic gradient descent (to accelerate convergence) are combined to drive the training process. The learning rate is flexibly adjusted based on the network's dynamic rhythm. Multiple rounds of iterations accurately refresh model weight parameters to improve predictive performance.

[0014] S7: Use multiple indicators to evaluate prediction results and compare them with the baseline. Develop a visualization interface to display links, node trajectories, topology evolution, and indicator trends. Optimize the model based on feedback to intuitively present and improve performance.

[0015] Preferably, the S1 step includes a multi-source data acquisition system, data preprocessing software, a network snapshot generation program and an adjacency matrix construction tool.

[0016] Preferably, the S2 step includes a feature encoding and fusion module, a dimensionality enhancement subsystem, a multi-head graph attention calculation engine and an embedding matrix generation software.

[0017] Preferably, the S3 step includes a GAT-LSTM network model, a hidden state splicing program, a multi-head self-attention allocation system, and a feature optimization and output module.

[0018] Preferably, the S4 step includes a multi-layer perceptron architecture, a hidden layer processing component and a Softmax conversion unit.

[0019] Preferably, the S5 step includes a loss function builder, an adaptive weighting strategy, an L2 regularization and label smoothing tool, and a cross-validation optimizer.

[0020] Preferably, the S6 step includes a data set partitioning tool, a hybrid optimization training system, a learning rate adaptive adjustment module and a model updating program.

[0021] Preferably, the S7 step includes a multi-index evaluation system, baseline comparison software, a visualization development tool and a model feedback optimizer.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. This invention uses a unique TGNN model architecture to accurately capture the spatiotemporal characteristics of nodes in opportunistic networks. The GAT in the embedding layer precisely analyzes node structural associations, while the LSTM and multi-head self-attention in the encoding layer collaboratively capture temporal evolution. The decoding layer efficiently maps features into a probabilistic space. This multi-mechanism fusion constructs a model with strong predictive capabilities, providing solid support for network communication planning, significantly enhancing the accuracy and stability of link predictions, and improving network communication reliability and data transmission success rates.

[0024] 2. This invention effectively addresses the challenges of opportunistic networks through innovative loss functions and optimization strategies. An adaptive weighting strategy addresses the problem of data sparsity. L2 regularization and label smoothing suppress overfitting and improve generalization. Multiple optimization algorithms collaborate to ensure the model's robustness and efficiency in dynamic and complex environments. The model can flexibly adapt to the characteristics of networks in multiple fields, demonstrating strong adaptability in scenarios such as intelligent transportation and disaster relief. It accurately predicts links based on scenario requirements, optimizes resource allocation and communication efficiency, and effectively promotes the development and innovation of network applications in multiple industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The overall process of the present invention Figure 1 ;

[0026] Figure 2 The process of the present invention Figure 2 ;

[0027] Figure 3 The process of the present invention Figure 3 ;

[0028] Figure 4 The process of the present invention Figure 4 ;

[0029] Figure 5 The process of the present invention Figure 5 ;

[0030] Figure 6 The process of the present invention Figure 6 ;

[0031] Figure 7 The process of the present invention Figure 7 ;

[0032] Figure 8 The process of the present invention Figure 8 . DETAILED DESCRIPTION

[0033] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0034] See Figure 1-8 , a method for predicting opportunistic network links based on a temporal graph neural network, comprising the following steps:

[0035] A method for predicting opportunistic network links based on a time-series graph neural network, comprising the following steps:

[0036] S1: Comprehensively collect multi-channel data within the opportunistic network, organize the data into a series of network snapshots according to predetermined time span rules, construct a weighted adjacency matrix to reflect the details of node connection interactions, and after normalization, lay a solid foundation for the underlying data architecture for link prediction.

[0037] S2: After accepting the weighted adjacency matrix and node feature materials, the encoding fusion operation is used to enhance the richness of feature dimensions. Attention calculation is used to accurately focus on the key features of the nodes. Weighted aggregation is used to generate an embedding matrix to capture the structural characteristics of the nodes and transfer the encoding process to deeply extract the feature connotations.

[0038] S3: Starting with node embedding materials, the spatial structure between nodes is mined with the help of graph attention networks, and the dynamic evolution trajectory of the node historical feature sequence is analyzed using long-short-term memory networks. The hidden states are spliced ​​and condensed into spatiotemporal embedding representations. The self-attention mechanism is used to dynamically adjust the weights according to the node coreness and time context, and the weights are transmitted to the decoding link after optimization.

[0039] S4: Start the multi-layer perceptron architecture, introduce the spatiotemporal embedding results of the coding layer into the input layer, use the deep abstract conversion function of the hidden layer, rely on the sophisticated processing of the activation function, and through multi-layer progressive conversion, map the features into a link generation probability matrix, clarify the probability of node emergence of the link, and provide a core decision-making reference for link prediction.

[0040] S5: Build a core architecture of the loss function around the predicted and real links, formulate a dynamic weighting strategy to address the link sparsity dilemma, integrate complexity management technology and label optimization methods, and use the cross-validation mechanism to fine-tune the weight parameters to achieve an ideal balance between model prediction accuracy and generalization breadth, thereby consolidating the model performance foundation.

[0041] S6: The dataset is divided into training, validation, and testing modules in appropriate proportions. Simulated annealing (to expand search breadth) and stochastic gradient descent (to accelerate convergence) are combined to drive the training process. The learning rate is flexibly adjusted based on the network's dynamic rhythm. Multiple rounds of iterations accurately refresh model weight parameters to improve predictive performance.

[0042] S7: Use multiple indicators to evaluate prediction results and compare them with the baseline. Develop a visualization interface to display links, node trajectories, topology evolution, and indicator trends. Optimize the model based on feedback to intuitively present and improve performance.

[0043] See Figure 2 The S1 step includes a multi-source data acquisition system, data preprocessing software, a network snapshot generation program and an adjacency matrix construction tool.

[0044] Specifically, the multi-source data acquisition system can collect data from multiple data sources such as sensors, communication nodes, logging devices, etc. in the opportunistic network, such as communication timestamps between nodes, signal strength, node geographic location coordinate information, node type or functional attribute identification, etc.

[0045] Furthermore, data preprocessing software can clean the collected data, remove obviously erroneous or abnormal data points, fill in missing data through interpolation or mean filling, and uniformly convert data in different formats into a format suitable for subsequent processing.

[0046] Furthermore, the network snapshot generation program slices the pre-processed data according to a preset time period (such as every 5 minutes or 10 minutes, etc.) to form network snapshots. Each snapshot represents the network status within the time period and records information such as the connection relationship and interaction frequency between nodes.

[0047] Furthermore, the adjacency matrix construction tool constructs the initial adjacency matrix based on the node connections in the network snapshot. If node i and node j are connected in the snapshot, the element at the corresponding position is set to 1, and if there is no connection, it is set to 0. t ={V,E t ,W t The weight W of the edge in t It can be expressed as the total duration of connections between nodes in the snapshot graph divided by the number of connections. The formula is as follows:

[0048]

[0049] in, f ij is the number of connections of node (i, j) on the tth network snapshot, ∑d ij is the total connection duration of node (i, j) on the tth network snapshot, and Δt represents the slice duration.

[0050] Implementation: First, a multi-source data collection system is activated to collect data from various devices and data sources within the opportunistic network. Once collected, the data is fed into data preprocessing software to ensure data quality and format consistency. Next, a network snapshot generation program is run to generate a sequence of network snapshots at a predetermined time granularity. Finally, an adjacency matrix construction tool is used to construct an initial adjacency matrix. Edge weights are calculated and normalized to obtain a weighted adjacency matrix, providing the data foundation for subsequent steps.

[0051] See Figure 3 The S2 step includes a feature encoding and fusion module, a dimension enhancement subsystem, a multi-head graph attention calculation engine and an embedding matrix generation software.

[0052] Specifically, the feature encoding and fusion module fuses the node's one-hot encoding (used to represent the node's unique identity) and position encoding (reflecting the node's location information in space) through a specific algorithm to generate an initial feature vector that can comprehensively represent the node's identity and position.

[0053] Furthermore, the dimensionality enhancement subsystem uses a learnable linear transformation matrix to process the initial feature vector, increasing the dimension of the feature vector so as to better capture the complex relationship between node features.

[0054] Furthermore, the multi-head graph attention calculation engine uses a multi-head graph attention mechanism to set up multiple attention heads. Each attention head focuses on a different subspace of the feature vector. The attention coefficient calculation formula can be referred to:

[0055]

[0056]

[0057] Among them, a T represents a shared attention mechanism, θ represents a shared linear transformation matrix, and then an activation function LeakyReLU function is used for activation. The above formula describes the importance of the feature of node j to the updated feature of node i. Finally, the softmax function is used to update e ij After normalization, n(i) represents the set of first-order neighbor nodes of node i.

[0058] Secondly, after introducing the multi-head attention mechanism, the calculation formula for calculating the average feature value between nodes is as follows:

[0059]

[0060] The model feature extraction ability and stable learning process are enriched by independently calculating and averaging multiple attention networks, and the attention coefficients are learned from multiple mapping subspaces to improve the model performance.

[0061] Further, the embedding matrix generation software can aggregate the attention coefficients calculated by each attention head and the corresponding node features to obtain the node embedding matrix, and the formula for mapping the attention coefficients to the snapshot adjacency matrix is as follows:

[0062]

[0063] Wherein, α ij The attention coefficients obtained by considering the multi-head attention mechanism can be obtained by each snapshot of the opportunity network. The dynamic opportunity network space feature can be represented by a set of three-dimensional variables, and the formula is as follows:

[0064]

[0065] Wherein, T is the number of snapshots, and N is the number of nodes in the opportunity network.

[0066] By automatically learning the relationship between nodes through the attention network, the attention coefficient matrix calculated can intuitively represent the spatial correlation between nodes, and the evolution of the attention coefficient matrix can be analyzed to observe the time sequence characteristics between nodes.

[0067] Implementation method: after receiving the weighted adjacency matrix and the initial feature vector of the node, the feature encoding and fusion module first performs encoding and fusion operation on the node features to generate the initial feature vector. Then, the dimension lifting subsystem performs dimension lifting processing on the vector. Then, the multi-head graph attention calculation engine calculates the attention coefficients between nodes according to the above formula, and multiple attention heads calculate the attention of different subspaces in parallel. Finally, the embedding matrix generation software aggregates the results of each attention head to generate the node embedding matrix, and transmits it to the encoding layer.

[0068] Referring to Figure 4 , the S3 step includes a GAT-LSTM network model, a hidden state splicing program, a multi-head self-attention distribution system, and a feature optimization and output module.

[0069] Specifically, the GAT-LSTM network model combines the advantages of the Graph Attention Network (GAT) and the Long Short-Term Memory Network (LSTM). When processing link prediction tasks in opportunistic networks, it works based on a weighted adjacency matrix and node feature vectors. It uses a multi-head graph attention mechanism to accurately calculate node attention coefficients. This mechanism can explore the complex structural relationships between nodes from multiple perspectives, and then performs a weighted aggregation operation on these coefficients to generate an embedding matrix. This embedding matrix accurately captures the structural characteristics of the nodes, providing a high-quality node representation foundation for subsequent LSTM processing.

[0070] Furthermore, the forward and reverse hidden states generated by the LSTM module in GAT-LSTM when processing the node embedding sequence are spliced ​​and integrated to achieve the convergence of multi-directional features. The formula for GAT-LSTM to extract the temporal features of network evolution and obtain the spatiotemporal embedding representation is as follows:

[0071]

[0072] Furthermore, the multi-head self-attention allocation system is based on the multi-head self-attention mechanism. It dynamically allocates attention weights to different feature dimensions of nodes according to the importance of the nodes in the network and their time dependency. The output definition of the multi-head self-attention mechanism is as follows:

[0073]

[0074] MultiHead(Q,K,V)=(head1;...;head i )

[0075] wherehead i =Attention(QW i Q ,KW i K ,VW i V )

[0076] Furthermore, the feature optimization and output module further optimizes the features after multi-head self-attention weighted update, such as feature compression or feature enhancement, to improve the expressiveness and effectiveness of the features, and then outputs the optimized spatiotemporal embedding representation to the decoding layer.

[0077] Implementation method: the node embedding matrix is input to start the GAT-LSTM network model processing process, GAT uses the weighted adjacency matrix and node feature vector to enable multi-head graph attention mechanism, carefully calculates the node attention coefficient, deeply mines the multi-subspace structure correlation, and weighted aggregation constructs the embedding matrix, which is used to deliver high-quality feature cornerstone. After receiving, the LSTM expands the bidirectional analysis of the node historical feature sequence, accurately extracts the forward and reverse hidden states, and splices them into a spatiotemporal embedding representation according to the rigorous formula, and then the multi-head self-attention distribution system calculates the attention coefficient according to the above formula and updates the features by weighting, and finally, the feature optimization and output module optimizes the weighted features and outputs to the decoding layer.

[0078] Referring to Figure 5 , the S4 step includes a multi-layer perceptron architecture, a hidden layer processing component, and a Softmax conversion unit.

[0079] Specifically, the multi-layer perceptron architecture is a neural network architecture composed of an input layer, a plurality of hidden layers, and an output layer. The input layer receives the spatiotemporal embedding representation from the encoding layer, and the hidden layers perform multi-layer feature abstraction and conversion, gradually extracting high-level features related to link prediction.

[0080] Further, the hidden layer processing component uses a rectified linear unit (ReLU) as an activation function. This activation function can increase the non-linear expression ability of the network, speed up the model training speed, and reduce the gradient vanishing problem. Through the stacking of multiple hidden layers, the input spatiotemporal embedding representation is deeply abstracted.

[0081] Further, the Softmax conversion unit is in the output layer of the multi-layer perceptron. It uses the Softmax function to convert the features output by the last hidden layer into a link generation probability matrix between node pairs.

[0082] The calculation method of each layer of the neural network is as follows:

[0083]

[0084]

[0085]

[0086] where h is the output result of the attention layer and They represent the weight matrix and bias coefficient matrix of the k-th layer neural network respectively. The activation function of the hidden layer adopts the sigmoid function, and the activation function of the output layer adopts the softmax function. The above formula clarifies the calculation method of each layer in the fully connected neural network structure, including the linear transformation from the input through the hidden layer to the output layer and the activation function processing to map the time series features to the probability space output link generation probability.

[0087] Implementation method: The input layer of the multi-layer perceptron architecture receives the spatiotemporal embedding representation from the encoding layer. The data enters the hidden layer processing component. Each neuron in the hidden layer first performs a linear transformation on the input, and then performs a nonlinear transformation through the ReLU activation function. After layer-by-layer processing through multiple hidden layers, higher-level features are extracted. Finally, the Softmax conversion unit converts the output of the last hidden layer into a link generation probability matrix through the Softmax function. This matrix represents the predicted probability of the existence of a link between each pair of nodes in the network.

[0088] See Figure 6 The S5 step includes a loss function builder, an adaptive weighting strategy, an L2 regularization and label smoothing tool, and a cross-validation optimizer.

[0089] Specifically, the loss function builder constructs a loss function based on the predicted link probability matrix and the actual link label to measure the difference between the model prediction results and the actual situation.

[0090] Furthermore, the adaptive weighting strategy, given the sparse links in opportunistic networks, sets dynamic weights for positive samples (samples with real links) to ensure accurate prediction. This increases their influence in the loss function based on sample characteristics (such as number ratio and prediction difficulty), thereby improving the model's sensitivity to positive examples.

[0091] Furthermore, L2 regularization and label smoothing tools: L2 regularization adds regularization terms to the loss function to constrain model complexity and prevent overfitting from causing generalization degradation; label smoothing fine-tunes the true label value, eases model confidence, and enhances generalization. The two work together to stabilize the model and improve performance.

[0092] Furthermore, the cross-validation optimizer uses cross-validation technology to split the dataset into multiple subsets, cyclically uses the subsets for validation and the remaining ones for training, optimizes the weight parameters based on the validation set indicators (loss value, accuracy, etc.), and uses search strategies (grid, random search) to find the optimal combination in the parameter domain to ensure good model generalization.

[0093] Among them, the formula for improving the loss function calculation method is as follows:

[0094]

[0095] Aiming at the sparsity of opportunity networks, a penalty value is introduced to improve the loss function based on the L2 norm. By processing the difference between the predicted value and the true value and setting the penalty value in different situations, the model bias is avoided to ensure the accuracy of zero-element prediction and overfitting.

[0096] For each training process, if Then p=1; if Then p=β>1, β is a positive value, which represents the penalty coefficient. This penalty matrix imposes a greater penalty on non-zero elements, thereby avoiding overfitting of the model to a certain extent. At the same time, this paper introduces the regularization term L based on the L2 norm. reg , the regularization loss is determined by calculating the sum of the squared norms of all weights in the model to ensure that the inputs between neural network layers have the same uniform distribution. Finally, a hybrid loss function is used to train and optimize the model, as shown in the following formula:

[0097] L total =L+αL reg

[0098] Among them, α is the proportion of regularization loss in the loss function.

[0099] Furthermore, an adaptive weighting strategy is used: in view of the sparsity of links in opportunistic networks, a dynamically increased weight is set for positive samples (i.e., samples with real links).

[0100] Implementation: The loss function builder constructs the core of the loss function based on the predicted and true link matrices, selecting methods such as cross entropy. An adaptive weighting strategy applies weights to the loss function based on the characteristics of positive examples. L2 regularization and label smoothing tools improve the loss function by adding regularization and smoothing labels, respectively. The cross-validation optimizer divides the dataset, performs multiple rounds of training and validation, and uses algorithms and search strategies such as gradient descent to find the optimal parameter set within the parameter range to reduce loss and improve generalization.

[0101] See Figure 7 The S6 step includes a data set partitioning tool, a hybrid optimization training system, a learning rate adaptive adjustment module and a model updating program.

[0102] Specifically, the dataset partitioning tool splits the opportunity network dataset into training set, validation set, and test set according to a specific ratio (such as 60% training, 20% validation, and 20% test), ensuring that the data distribution reflects the overall characteristics, laying the foundation for model evaluation, preventing overfitting and underfitting, and balancing model learning and evaluation.

[0103] Furthermore, the hybrid optimization training system integrates simulated annealing and stochastic gradient descent algorithms to optimize training. Simulated annealing allows acceptance of inferior solutions in the initial stage to avoid local optimality and expand the search domain; stochastic gradient descent uses sample gradients to quickly update parameters to promote convergence, and collaborates during iteration. Stochastic gradient descent is first used to obtain gradient updates, and then simulated annealing is used to determine acceptance, thereby improving training efficiency and effectiveness.

[0104] Furthermore, the learning rate adaptive adjustment module dynamically adjusts the learning rate based on node mobility (speed and direction changes) and network dynamics (topology frequency changes and link duration distribution). Increasing the learning rate during periods of network turbulence accelerates model adaptation; decreasing the learning rate during periods of stability stabilizes training and fine-tunes parameters, ensuring the model accurately tracks network characteristics.

[0105] The model update procedure training iterations updates the model weight parameters according to the gradient and learning rate.

[0106] Implementation method: The dataset partitioning tool divides the dataset proportionally to prepare data for model training and evaluation. The hybrid optimization training system uses algorithms to collaboratively update parameters in the training loop. The learning rate adaptive adjustment module monitors network characteristics and calculates the learning rate. The model update program iteratively updates the weights using gradients and learning rates. Multiple rounds of optimization are carried out until the model performance steadily improves and adapts to network dynamics.

[0107] See Figure 8 The S7 step includes a multi-index evaluation system, baseline comparison software, visualization development tools and model feedback optimizer.

[0108] Specifically, the multi-index evaluation system uses multiple indicators such as accuracy, recall rate, F1 value, mean square error and AUC-ROC curve to quantitatively evaluate the model link prediction performance from the dimensions of prediction accuracy, comprehensiveness, error and classification performance, and analyze the advantages and disadvantages and generalization ability of the model from multiple angles.

[0109] Furthermore, the baseline comparison software compares the prediction results of this model with those of baseline models (traditional statistics, simple machine learning models), explains the advantages and disadvantages of the model based on indicator values ​​(accuracy, recall rate difference, etc.), verifies its advancedness in opportunistic network link prediction, and helps researchers identify optimization directions.

[0110] Furthermore, visualization development tools create a visualization interface that uses graphics (node-link diagrams, trajectory diagrams, topology evolution animations, and indicator trend line charts) to intuitively present predicted links, node trajectories, topology changes, and indicator trends, helping operation and maintenance personnel gain insight into network performance and prediction status, and providing a visualization basis for decision optimization.

[0111] Furthermore, the model feedback optimizer deeply explores the deficiencies of the model based on visual feedback and evaluation results. For example, if the link prediction in a certain area is poor, it will analyze and process the data and adjust the parameters in detail. It will optimize the training strategy (variable learning rate strategy, data set ratio) according to the indicator trend, and cyclically optimize to improve the model performance and adaptability, thereby achieving accurate link prediction.

[0112] Implementation method: The multi-indicator evaluation system calculates the indicator values ​​of the model test set, the baseline comparison software compares the model with the baseline indicators to find the gap, the visualization development tool visualizes the prediction, node, topology, and indicator information, and the model feedback optimizer focuses on the problem areas based on the feedback results and adjusts the model data processing, parameters, and training strategies according to the indicator trends. Multiple rounds of cycles are used to improve performance and adapt to the network link prediction requirements.

[0113] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting opportunistic network links based on a time-series graph neural network, comprising the following steps: S1: Collect multi-channel data within the opportunistic network, organize the data into a series of network snapshots according to predetermined time span rules, construct a matrix to reflect the details of node connection interactions, and normalize the data to lay the foundation for the underlying data architecture for link prediction; S2: After receiving the weighted adjacency matrix and node feature materials, the encoding fusion operation is performed to enhance the feature dimension richness. Attention calculation is used to accurately focus on the key features of the node. The weighted aggregation generates an embedding matrix to capture the node structural characteristics and transfer the encoding process. S3: Using node embedding materials as a starting point, the graph attention network is used to mine the spatial structure between nodes. The long short-term memory network is used to analyze the dynamic evolution trajectory of the node's historical feature sequence. The hidden states are spliced ​​to form a spatiotemporal embedding representation. The self-attention mechanism is used to dynamically allocate weights based on the node's coreness and time context. After optimization, it is transmitted to the decoding stage. S4: The multi-layer perceptron architecture is started. The input layer introduces the spatiotemporal embedding results of the encoding layer. Based on the hidden layer's deep abstract transformation function and relying on the activation function processing, the features are mapped into a link generation probability matrix through progressive transformation. This clarifies the probability of link occurrence between node pairs and provides a core decision-making reference for link prediction. S5: Build a core loss function architecture around predicted and real links, develop a dynamic weighting strategy to address link sparsity, integrate complexity management techniques with label optimization methods, and optimize weight parameters through a cross-validation mechanism to stabilize model performance. S6: Split the dataset into training, validation, and test sets, integrate simulated annealing to expand the search breadth and stochastic gradient descent to accelerate the convergence training process, flexibly adjust the learning rate based on network dynamics, and iteratively refresh the model weight parameters; S7: Use multiple indicators to evaluate prediction results and compare them with the baseline. Develop a visualization interface to display links, node trajectories, topology evolution, and indicator trends. Optimize the model based on feedback to intuitively present and improve performance.

2. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S1 step includes a multi-source data acquisition system, data preprocessing software, a network snapshot generation program and an adjacency matrix construction tool.

3. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S2 step includes a feature encoding and fusion module, a dimensionality enhancement subsystem, a multi-head graph attention calculation engine, and an embedding matrix generation software.

4. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S3 step includes a GAT-LSTM network model, a hidden state concatenation program, a multi-head self-attention allocation system, and a feature optimization and output module.

5. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S4 step includes a multi-layer perceptron architecture, a hidden layer processing component, and a Softmax conversion unit.

6. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S5 step includes a loss function builder, an adaptive weighting strategy, an L2 regularization and label smoothing tool, and a cross-validation optimizer.

7. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S6 step includes a data set partitioning tool, a hybrid optimization training system, a learning rate adaptive adjustment module and a model updating program.

8. The method for predicting opportunistic network links based on a temporal graph neural network according to claim 1, characterized in that: The S7 step includes a multi-index evaluation system, baseline comparison software, visualization development tools and model feedback optimizer.

Images