Enhancing Cross-Domain Graph Classification with Multi-Teacher Knowledge Distillation

Graph classification, as one of the core issues in graph data analysis, demonstrates significant application value in scenarios such as chemical molecular property prediction and protein function analysis. Given that graph data contains both domain-specific feature knowledge and structural knowledge with lower domain dependency, transferring and integrating structural knowledge across heterogeneous graph models from different domains is a viable approach to enhancing graph classification performance in each domain. Motivated by this, we propose a multi-teacher knowledge distillation framework tailored for heterogeneous cross-domain graph data. Under the federated learning setting, this framework effectively utilizes and transfers the structural knowledge embedded in different domain-specific graph models through knowledge distillation. Specifically, the framework employs a knowledge separation model to separate structural knowledge from feature knowledge, where the structural knowledge module takes domain-independent node structural vectors as input, while the feature knowledge module retains local personalized training. Furthermore, the framework incorporates a multi-teacher knowledge distillation mechanism within the federated graph learning setting to fully transfer structural features. Experimental results demonstrate that the proposed framework outperforms existing methods in various cross-domain settings for graph classification.