A Graph Representation Learning Algorithm Based on Attention Mechanism and Node Similarity

Abstract

Recently graph representation learning has attracted much attention of researchers, aiming to capture and preserve the graph structure by encoding it into low-dimensional vectors. Attention mechanism is a recent research hotspot in learning the representation of graph. In this paper, a graph representation learning algorithm based on Attention Mechanism and Node Similarity (AMNS for short) is proposed. Firstly, the similarity neighborhood is generated for each node in graph. Secondly, attention mechanism is used to learn weight coefficients for each node and its similarity neighborhood. Thirdly, the node vectors are generated by aggregating its similarity neighborhood with weight coefficients. Finally, node vectors are applied to many tasks, e.g., node classification and clustering. The experiments on real-world network datasets prove that the AMNS algorithm achieves excellent results.