CN-121984877-A - Multi-hop link prediction method, system, equipment and medium based on federal comparison learning

CN121984877ACN 121984877 ACN121984877 ACN 121984877ACN-121984877-A

Abstract

The invention discloses a multi-hop link prediction method, a system, equipment and a medium based on federal comparison learning, which belong to the technical field of power communication link prediction and comprise the steps of constructing a pruning sub-graph for a target link in a local private graph data set and endowing the pruning sub-graph with edge attributes, running a graph attention network, generating a node representation of structural perception of a target link corresponding to a pruning line graph node, calculating cross entropy loss and contrast loss, constructing a local total loss function, training a local map neural network model by taking the minimized local total loss function as a target to obtain updated model parameters, calculating parameter updating quantity and aggregated model parameters, and applying the aggregated model parameters to a multi-hop link prediction model to predict to generate a prediction result.

Inventors

JIANG XUEJIAO
WU HAIJIE
ZHONG LEI
CHEN LINCONG
XU JIALONG
WU MIN

Assignees

海南电网有限责任公司

Dates

Publication Date: 20260505
Application Date: 20251210

Claims (10)

1. A multi-hop link prediction method based on federal contrast learning is characterized by comprising the following steps of, Constructing a local private graph data set in a federal learning environment, and constructing a pruning sub-graph for a target link in the local private graph data set; Converting the pruning subgraph into a pruning line graph and endowing the pruning line graph with edge attributes; The method comprises the steps of (1) operating a graph attention network on a pruning line graph to generate a node representation of structural perception of a target link corresponding to a pruning line graph node; calculating cross entropy loss according to the node representation of the structure perception, constructing a comparison learning target, calculating comparison loss, and constructing a local total loss function according to the cross entropy loss and the comparison loss; Each client updates model parameters on the local data with the minimum local total loss function as a target and calculates parameter updating quantity; And calculating parameter differences and weighting coefficients according to the parameter updating quantity, calculating aggregate model parameters according to the weighting coefficients and the updating model parameters, and applying the aggregate model parameters to a multi-hop link prediction model for prediction to generate a prediction result.
2. The multi-hop link prediction method based on federal contrast learning of claim 1, wherein the steps of constructing a local private graph dataset in a federal learning environment and constructing a pruned subgraph for a target link in the local private graph dataset comprise: Deploying a federal learning system comprising a set number of clients, each client holding a local private graph dataset; Selecting a node pair to be predicted from a local private graph data set as a target link; extracting a closed subgraph from a local private graph dataset by taking two endpoints of a target link as centers; pruning the closed subgraph according to the node attribute similarity to generate a pruned subgraph.
3. The multi-hop link prediction method based on federal contrast learning according to claim 2, wherein the step of converting the pruned subgraph into a pruned line graph and assigning edge attributes to the pruned line graph comprises: Setting each side of the pruning subgraph as a node in the line graph, and generating a pruning line graph; an edge attribute is assigned to each node in the pruned line graph.
4. A multi-hop link prediction method based on federal contrast learning as claimed in claim 3, wherein the step of generating a structurally perceived node representation of the target link corresponding to the pruning line graph node by the running graph attention network on the pruning line graph comprises: Determining a target node of a target link in a pruning line graph; Calculating the attention coefficients of a target node and a neighbor node in the pruning line graph; and running a plurality of attention heads in parallel, respectively calculating the attention coefficients of all the attention heads, and then averaging to obtain the node representation of structural perception.
5. The federal contrast learning-based multi-hop link prediction method according to claim 4, wherein the steps of calculating cross entropy loss from the node representation of the structure perception, constructing a contrast learning objective and calculating contrast loss, and constructing a local total loss function from the cross entropy loss and the contrast loss comprise: calculating two kinds of cross entropy loss for link prediction according to the node representation perceived by the structure and the real label of the target link; the node representation of structural perception generated by the target link is used as a positive sample, and Gaussian noise disturbance is applied to the node representation of structural perception as a negative sample to construct a comparison learning target; Calculating contrast loss according to the contrast learning target; And constructing a local total loss function according to the cross entropy loss and the contrast loss.
6. The federal contrast learning-based multi-hop link prediction method according to claim 5, wherein the step of updating the model parameters and calculating the parameter update amount with each client targeting to minimize the local total loss function on the local data comprises: Before starting local training, each client receives global model parameters of the current round from a server as local model parameters; Updating the local model parameters according to the local total loss function to obtain updated model parameters; and calculating the parameter updating quantity according to the updated model parameters and the local model parameters before updating.
7. The method for multi-hop link prediction based on federal contrast learning of claim 6, wherein the step of calculating the parameter difference and the weighting coefficient according to the parameter update amount, calculating the aggregate model parameter according to the weighting coefficient and the update model parameter, and applying the aggregate model parameter to the multi-hop link prediction model to perform prediction, and generating the prediction result comprises: calculating a parameter difference according to the parameter updating amount; calculating a weighting coefficient according to the parameter difference; calculating aggregate model parameters according to the weighting coefficients and the updated model parameters; and applying the aggregation model parameters to a multi-hop link prediction model to predict, and generating a prediction result.
8. A multi-hop link prediction system based on federal contrast learning, applying a multi-hop link prediction method based on federal contrast learning as claimed in any one of claims 1 to 7, comprising: the data set construction module is used for constructing a local private graph data set in the federation learning environment and constructing a pruning sub-graph for a target link in the local private graph data set; The line diagram conversion module is used for converting the pruning sub-graph into a pruning line diagram and endowing the pruning line diagram with edge attributes; the diagram attention module is used for generating a node representation of structural perception of a target link corresponding to a pruning line diagram node by running a diagram attention network on the pruning line diagram; The loss calculation module is used for calculating cross entropy loss according to the node representation of the structure perception, constructing a comparison learning target, calculating comparison loss and constructing a local total loss function according to the cross entropy loss and the comparison loss; The parameter updating module is used for updating the model parameters on the local data by taking the minimum local total loss function as a target and calculating the parameter updating quantity; And the parameter aggregation module is used for calculating parameter differences and weighting coefficients according to the parameter updating quantity, calculating an aggregation model parameter according to the weighting coefficients and the updating model parameter, and applying the aggregation model parameter to the multi-hop link prediction model for prediction to generate a prediction result.
9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of a federal contrast learning-based multi-hop link prediction method according to any one of claims 1 to 7 when the computer program is executed.
10. A computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the steps of a federal contrast learning based multi-hop link prediction method according to any of claims 1 to 7.

Description

Multi-hop link prediction method, system, equipment and medium based on federal comparison learning Technical Field The invention relates to the technical field of power communication link prediction, in particular to a multi-hop link prediction method, a multi-hop link prediction system, multi-hop link prediction equipment and a multi-hop link prediction medium based on federal comparison learning. Background Link prediction is a basic task of network analysis, solves the key challenge of predicting potential connection between nodes, and has wide application in various fields. In general, this research area has undergone the evolution of three major paradigms, each providing different advantages and limitations. The initial paradigm employs heuristic-based similarity metrics to directly quantify node relationships through topological patterns. DeepWalk and node2vec, etc., have introduced matrix decomposition techniques to construct low-dimensional representations, providing finer relational modeling. Although these methods utilize network structures, content-based methods integrate explicit node properties, thereby improving prediction accuracy. However, one challenge for heuristic approaches is their reliance on artificial feature engineering of the network. The prior art overcomes this limitation by combining a hybrid approach of topology and node features, such as the SEAL framework, which is a GNN-based model in which predicted features can be learned from neighborhood, in which innovations a Graphic Automatic Encoder (GAE) represents a flexible architecture that generates node embeddings through the GNN layer and then aggregates the representations to form link-specific features, thereby enabling adaptive feature extraction for different graphic features. With industry growing concern for data privacy protection, certain areas of communication links and device conditions constitute private data, which makes Federal Learning (FL) an effective solution in this regard. The federal learning is used as a distributed machine learning strategy for privacy sensitive data analysis, an effective machine learning framework can be constructed, and the data analysis capability under privacy constraint is enhanced. Recent advances in federal learning have changed privacy preserving link prediction techniques across distributed networks. The existing method mainly adopts three architecture examples, namely a GRAPHSAGE and other graphic neural networks are expanded to federal settings through embedded federal learning, local topology privacy is kept, a hybrid federal learning framework combining differential privacy and a graph self-encoder is used for balancing prediction precision and privacy guarantee, and a dynamic federal learning system adopts an evolution network of time random walk and incremental learning. Recent advances in graph neural networks have established federal learning as a class of mainstream approaches to link prediction tasks, with the graph neural networks providing a breakthrough solution to the characteristic over-average problem that is unique to conventional GNN architectures. However, the line graph approach is still an unresolved challenge to accommodate privacy preserving frameworks, where links exhibit both local and global dependencies in the power system wireless network, while the node neighborhood provides the necessary information, and the collective configuration of the full network links severely impacts the connectivity pattern, making pure local analysis inadequate. Furthermore, regional topological heterogeneity introduces cross-regional divergent links, which in turn leads to significant parameter differences between the locally trained models. This divergence propagates to the federal aggregation stage, where the global model becomes biased towards certain regional topologies, ultimately reducing the accuracy of predictions of network configurations that are not representative. Disclosure of Invention The present invention has been made in view of the above-described problems. The technical problem solved by the invention is that in a wireless network of an electric power system, link formation simultaneously shows local and global dependence, a node neighborhood provides necessary information, the collective configuration of the whole network links seriously influences a connection mode, so that pure local analysis is insufficient, regional topology heterogeneity introduces the distribution of regional branched links, which in turn leads to the parameter difference among local training models, the branching propagates to a federal aggregation stage, the global model becomes biased to certain regional topologies, and the prediction precision of the network configuration with representative deficiency is reduced. In order to solve the technical problems, the invention provides a multi-hop link prediction method based on federal contrast learning, which comprises the followi