HashAlign: Hash-based alignment of multiple graphs

Abstract

Fusing or aligning two or more networks is a fundamental building block of many graph mining tasks (e.g., recommendation systems, link prediction, collective analysis of networks). Most past work has focused on formulating pairwise graph alignment as an optimization problem with varying constraints and relaxations. In this paper, we study the problem of multiple graph alignment (collectively aligning multiple graphs at once) and propose HashAlign, an efficient and intuitive hash-based framework for network alignment that leverages structural properties and other node and edge attributes (if available) simultaneously. We introduce a new construction of LSH families, as well as robust node and graph features that are tailored for this task. Our method quickly aligns multiple graphs while avoiding the all-pairwise-comparison problem by expressing all alignments in terms of a chosen ‘center’ graph. Our extensive experiments on synthetic and real networks show that, on average, HashAlign is 2 × faster and 10 to 20% more accurate than the baselines in pairwise alignment, and 2 × faster while 50% more accurate in multiple graph alignment.