Social Network Classification Based on Adaptive Structure-Alignment Federated Graph Learning Framework

Federated learning plays a crucial role in social network classification. By performing model training on local devices and sharing only parameter updates, it not only enhances the generalization ability of the model, but also provides support for personalized recommendation systems and protects the data privacy of the local clients. However, local clients usually have incomplete graphs and misrepresent the graph structure. Structural heterogeneity poses challenges for the model in effectively learning significant similarity graph patterns among different clients. Existing federated graph learning methods have limitations and cannot handle Non-IID graph data well, most federated graph learning methods aggregate all learnable parameters, complicating the extraction and sharing of specific structural knowledge among clients. Therefore, this paper proposes an adaptive structure-alignment federated graph learning framework(ASAFGL), which utilizes a structure encoder to separate structural information from heterogeneous graph data and realize federated knowledge sharing during local training. The structure encoder is uniformly shared among clients, whereas the feature-based knowledge is acquired in a personalized manner. We conduct comprehensive experiments across various social network datasets and in cross-dataset Non-IID settings, showcasing the superiority of ASAFGL. Furthermore, the results also show that the framework has good cross-domain generalization ability.