Graph neural networks with multiple kernel ensemble attention

Abstract

Graph neural networks have recently attracted increasing research attention. Recent research has shown that applying the attention mechanism is helpful for graph learning. The performance of the attention function is critical for graph learning. In this paper, we propose a multiple kernel ensemble attention method for graph learning. Unlike previous work, the attention weights in the proposed method are determined by comparing the similarity between features in the reproducing kernel Hilbert space. Our network can automatically learn the optimal kernel function from a set of predefined candidate kernels which are of the same type. This significantly eases the hyperparameter tuning procedure. We show that our network is effective to improve the graph learning performance. To the best of our knowledge, this is the first work that incorporates automatic multiple kernel learning into graph convolutional networks. We validate our graph network for both transductive learning and inductive learning on four benchmark datasets: Cora, Citesser, Pubmed, and the protein–protein interaction dataset. The experimental results show that we achieve competitive performance when compared with the latest work.