Decision-based Sampling for Node Context Representation

Abstract

Learning low dimensional representations requires an expressive technique capable of capturing the different features for nodes, the relationship between nodes in the network and thus their similarities. However, many existing embedding techniques focus only on capturing the structural patterns in the network by randomly sampling the nodes in the neighborhood of the target node. To deal with this issue, we propose DSNCR, a node representation framework which uses the non-linear node attributes as well as their neighbourhood structural information to capture nodes similarities. This approach computes a semi-supervised regression analysis on the node attributes to guide a flexible probability walk procedure, such that different neighbourhoods are explored to capture rich network attributes and structures in a learned embedding. We verify the effectiveness of our model on link prediction and node classification tasks using real-life benchmark datasets, for which our technique performs better than existing embedding methods.