Concept-cognitive Learning-empowered Federated Graph learning paradigm for social networks

Graph Convolution Networks (GCNs) are methods that can extract features from graph data and analyze graph data. Specially, GCNs plays an important role in the realm of social network analysis. In social networks, the arrival of new users and the formation of new social ties have led to the continuous growth of the graph data to be processed by GCNs, and the secure exchange of graph data stored in different terminals has become a major problem. Therefore, GCNs faces the following two key problems: 1) insufficient vertex feature extraction from large graph data leads to low accuracy of GCNs training results and high training overhead of GCNs training process; 2) insufficient data integrity and user privacy cannot be guaranteed in model training. To tackle these problems, we propose a Concept-cognitive Learning-empowered Federated Graph Learning (CCL-FedGCN) framework, which uses the Formal Concept Analysis (FCA) to fully mine graph data and improve the training accuracy of GCNs models. Under the federated learning (FL), GCNs models can be trained independently in different clients, and the local models can be optimized by sharing model parameters, which ensures the integrity of the data and improves the training accuracy of the GCNs model under the premise of protecting data privacy. Compared with multiple baselines on real-world multi-type datasets, CCL-FedGCN has better performance on vertex classification, accuracy improvements range from 0.6% to 7.02%.