Network embedding aims to learn a vector for each node while preserves inherent properties of the network. Topological structure and node attributes are both critical for understanding the network formulation. This paper focuses on making the topological and attributed properties complement each other. We propose FATNet for integrating the global relations of the topology and attributes into robust representations. Specifically, non-local attribute relations are proposed to capture long-distance dependencies for enriching the topological structure. Meanwhile, we design attributes smoothing filter to preserve the critical attribute values, while interpolating global topology relations via high-order proximity. These relations provide reasonable principles to fuse the structure and attribute for network embedding. Extensive experiments are carried out with five real-world datasets on four downstream tasks, including node classification, link prediction, node clustering and graph visualization. Experiments have shown that the FATNet can achieve superior performance in most cases.

Top