Dynamic-n-Static Multiplex Graph Representation Learning for Improved Link Prediction.

Abstract

Social systems often involve multiple types of relations, each exhibiting distinct temporal characteristics. Such systems can be modeled as temporal multiplex graphs in which each graph layer represents one type of relation. In this article, we study the link prediction problem in a particular yet common class of multiplex graphs, namely dynamic-n-static multiplex graphs, which consist of both evolving dynamic layers and long-term static layers. We propose DS-MGN, a representation learning model for dynamic-n-static multiplex graphs. DS-MGN introduces a cross-layer neighbor encoding (CLNE) scheme, which injects stable topological priors from static layers as contextual information, thereby enhancing the representations of the more complex dynamic layers by encoding nodes' structural connectivity patterns in the static layer. The CLNE scheme is architecture-compatible and can be seamlessly integrated into other models with similar neural network architectures, offering substantial performance improvements. Furthermore, to support research on such complex graph structures, we present a high-quality dynamic-n-static multiplex graph dataset, which includes a dynamic sports social network as well as a static network capturing long-term stable collegial relationships between members, meeting the need for complex and comprehensive multiplex graph datasets under practical data collection challenges. Extensive experiments on both a public benchmark and the collected dataset demonstrate that DS-MGN achieves state-of-the-art performance against 12 strong baselines. Moreover, our analyses highlight its scalability and potential for real-world applications, including evolving social network analysis, professional collaboration tracking, and dynamic recommendation systems.