Cross-Temporal Snapshot Alignment for Dynamic Multi-Relational Networks

Abstract

A dynamic network is often represented as a sequence of snapshots evolving over time. In certain real-world scenarios, the identities of nodes in snapshots of a dynamic network are unknown and need to be figured out. To deal with such a challenge, recently, the task of cross-temporal snapshot alignment for dynamic networks is proposed, which aims to match equivalent nodes across temporal snapshots of a dynamic network given a small set of identified nodes. However, in many dynamic multi-relational networks like temporal knowledge graphs, the relation type information of edges, which can be useful for the alignment task, is neglected by existing methods. In this paper, we focus on a similar task but pay special attention to dynamic multi-relational networks. We propose a Relation-Aware Cross-Temporal Snapshot Alignment model (RCTSA) that incorporates time-dependent topological structure information into temporal node embeddings and temporal relation embeddings. To disentangle time-dependent information and time-independent information of the dynamic multi-relational network, RCTSA maintains a time embedding for each snapshot to preserve the temporal information which is incorporated into entity embeddings and relation embeddings to get the temporal embeddings. After training, the learned entity embeddings of unidentified nodes together with time embeddings can be used for the alignment task. Experimental results on real-world dynamic multi-relational networks demonstrate the effectiveness of our model.