HAMF: Highly Adaptable Multi-graph Fusion for Cross-domain Graphs

Nowadays, many realistic networks are modeled as multiple graphs that interact with each other. Therefore, multi-graph fusion comes into being. Most existing algorithms are oriented to a particular scenario, which lacks universality and thus can not be applied to other scenarios with varying structures. Besides, the extraction of node information is insufficient, and the fusion of cross-graph relationships needs to be better considered. To solve these problems, we propose a Highly Adaptable Multi-graph Fusion (HAMF) model for cross-domain graphs, a universal framework that can handle different graph structures, including bipartite, homogeneous, and heterogeneous graphs. In addition, we use interactive information to guide the feature extraction at the node level and use a gate mechanism to measure the relationship at the graph level, after which we make predictions accordingly. Experiments are conducted on four real datasets involving recommendation systems, citation networks, and movie networks. The results show that the predicting error decreases up to 30% compared to the baselines.