Data Representation via Attribute Selection- Propagation Neural Network

Abstract

Graph Convolutional Networks (GCNs) have been commonly studied for attribute graph data representation. It is known that the core of Graph Convolution (GC) operation is to define a specific graph propagation operation for graph node's attributes. Existing GCs mainly perform propagation over node's all attributes. However, this ‘full’ propagation may be unreasonable on some practical learning tasks and also be vulnerable to the noisy attributes. To address this issue, in this paper, we propose a novel Attribute Selection-Propagation (ASP) mechanism for attribute graph data representation by incorporating attribute selection into GCs. The main aspect of the proposed ASP is that it can be formulated as a <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">regularization model</b> based on which we can derive a simple update rule to implement ASP in a <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">self-supervised</b> manner. ASP aims to adaptively propagate some optimal attributes of node to better serve message passing in GCs. Using ASP, we then present a novel graph neural network, named ASPNet for attribute graph representation and learning. Experiments on several graph learning tasks including node classification, clustering and link prediction demonstrate the effectiveness of the proposed ASPNet.