Graph neural networks have been attracted attention in recent years due to their superior performance in many tasks including graph classification. Unlike deep networks for image data, in which deep networks can work directly on raw data, graph neural networks usually require to be thoughtfully designed to integrate topological information (via adjacency matrices) to the networks, rendering high-computational cost and complicating the networks. This paper introduces a novel framework, which first employs an effective process to convert adjacency matrices into more convenient Column Ordering Free (COF) matrices and then builds a deep network to learn high-level representations from these matrices for graph classification. Our framework is not only permutation invariant but also has computational efficiency. In particular, we show that our network achieves comparable performance with other existing graph neural networks but at least 3 times faster in speed on a set of both biological and social network benchmark datasets.
Institute of Electrical and Electronics Engineers Inc.