A-GNN: Anchors-Aware Graph Neural Networks for Node Embedding

Abstract

With the rapid development of information technology, it has become increasingly popular to handle and analyze complex relationships of various information network applications, such as social networks and biological networks. An unsolved primary challenge is to find a way to represent the network structure to efficiently compute, process and analyze network tasks. Graph Neural Network (GNN) based node representation learning is an emerging learning paradigm that embeds network nodes into a low dimensional vector space through preserving the network topology as possible. However, existing GNN architectures have limitation in distinguishing the position of nodes with the similar topology, which is crucial for many network prediction and classification tasks. Anchors are defined as special nodes which are in the important positions, and carries a lot of interactive information with other normal nodes. In this paper, we propose Anchors-aware Graph Neural Networks (A-GNN), which can make the vectors of node embedding contain location information by introducing anchors. A-GNN first selects the set of anchors, computes the distance of any given target node to each anchor, and afterwards learns a non-linear distance-weighted aggregation scheme over the anchors. Therefore A-GNN can obtain global position information of nodes regarding the anchors. A-GNN are applied to multiple prediction tasks including link prediction and node classification. Experimental results show that our model is superior to other GNN architectures on six datasets, in terms of the ROC, AUC accuracy score.