Position and structure-aware graph learning

Abstract

Graph representation learning methods have gained great popularity in tackling various analytics tasks, such as node classification, link prediction and graph classification. However, most of graph representation learning methods merely consider the local information around the nodes. In addition, information about the global relative position information of each node is not fully taken into account. Recently, Position-aware Graph Neural Networks have been proposed to capture positions of nodes with respect to the anchor nodes and embed this position information into the feature representation, which aims to distinguish nodes which are locally structurally similar. This network aggregates position information from the randomly sampled anchor set nodes to the given target node. With such randomness, the sampled nodes lack particularity, and not all of them are significant enough in the graph to be used as position reference points. To address these issues, we propose a Position and Structure-aware Graph Learning framework (PSGL). The proposed PSGL extracts local topology information of nodes through the structure representation module, obtains anchor set nodes through graph pooling and further calculates global position information of the nodes to each anchor set node. The attention mechanism is used to weigh the obtained local topology and the global position information adaptively. Our proposed PSGL is capable of learning structural and positional information adaptively and encoding more informative characteristics in real-world networks. Our experimental results demonstrate that the proposed PSGL outperforms the state-of-the-art graph representation learning methods. Our code is available at https://github.com/leaf-ygq/PSGL.