Non-IID Transfer Learning on Graphs

Transfer learning refers to the transfer of knowledge or information from a relevant source domain target domain. However, most existing theories and algorithms focus on IID tasks, where source/target samples are assumed be independent identically distributed. Very little effort is devoted theoretically studying transferability non-IID e.g., cross-network mining. To bridge gap, in this paper, we propose rigorous generalization bounds for graph graph. The crucial idea characterize perspective Weisfeiler-Lehman isomorphism test. end, novel Graph Subtree Discrepancy measure distribution shift between graphs. Then error learning, including both node classification link prediction can derived terms across domains. This thereby motivates us generic adaptive network (GRADE) minimize graphs learning. Experimental results verify effectiveness efficiency our GRADE framework cross-domain recommendation tasks.