First and Complementary Neighborhood Combination of Adjacency Matrix for Graph Learning.

A variety of effective frameworks for representation learning on graphs have been recently developed, such as the graph neural networks and the WL-kernel based models. These frameworks rely on complex modules with heavy computation and involve lots of parameters. In contrast, the transformation of adjacency matrix method is computationally simple and could improve performance with a shorter learning process. In this paper, we introduce the notion of complementary neighborhood, and propose to combine the first and the complementary neighborhood of the adjacency matrix for graph learning. The proposed transformation can be applied to several state-of-the-art graph models. Based on a series of experiments conducted on several benchmark datasets, we show that the proposed approach improves the state-of-the-art performance for four kinds of graph tasks, including supervised and semi-supervised graph classification, graph link prediction, graph edge generation and classification. The implementation codes are available online at \url{https://github.com/CODE-SUBMIT/Graph_Neighborhood_1}.