Network Embedding with Attribute Refinement

Presented at the 16th International Workshop on Mining and Learning with Graphs (MLG), co-located with KDD 2020.

Abstract

Network embedding has been an active research area given its effectiveness in mapping nodes to low-dimensional representations. While previous studies mostly focus on network topology, recent
advances have shown that rich node-level information, known as attributes, often exist and can substantially benefit embedding based on the assumption of homophily: nodes often connect to other nodes similar to themselves. However, we find that inconsistencies often occur in node attributes in real-world data, which can undermine the homophily assumption and thus degrade the performance of attributed network embedding. To address this drawback, in this paper, we present a novel framework for unsupervised network embedding with attribute refinement. In particular, we propose a learnable filter to automatically refine the individual attributes of every node. To overcome the challenge of no supervision, we leverage homophily to guide the refinement—attributes should be fine-tuned in a way to reinforce the correlation with topology. Finally, we conduct extensive experiments on three benchmark real-world datasets, which show that our model significantly outperforms state-of-the-art methods on both node classification and link prediction tasks. Furthermore, we perform model analysis to demonstrate that our framework can effectively refine attributes.

PDF Link: http://www.mlgworkshop.org/2020/papers/MLG2020_paper_19.pdf