Representation learning in multiplex graphs: Where and how to fuse information?
Piotr Bielak, Tomasz Kajdanowicz
TL;DR
This paper tackles node representation learning in multiplex graphs with multiple edge types under unsupervised and self-supervised settings. It introduces a taxonomy of information fusion locations along the graph-learning pipeline—graph-level, GNN-level, embedding-level, and prediction-level—and provides an extensive experimental evaluation across six real-world datasets. The authors compare a wide range of fusion schemes and propose extensions (e.g., F(GBT,*) and F(DGI,*)) to improve performance on node classification, clustering, and similarity search, highlighting the strengths of GNN-level fusion (DMGI/HDGI) and the trade-offs of embedding- and prediction-level methods. The study also discusses inductivity constraints, limitations of lookup-based fusion, and future directions for designing dedicated multiplex GNN architectures and more scalable fusion mechanisms.
Abstract
In recent years, unsupervised and self-supervised graph representation learning has gained popularity in the research community. However, most proposed methods are focused on homogeneous networks, whereas real-world graphs often contain multiple node and edge types. Multiplex graphs, a special type of heterogeneous graphs, possess richer information, provide better modeling capabilities and integrate more detailed data from potentially different sources. The diverse edge types in multiplex graphs provide more context and insights into the underlying processes of representation learning. In this paper, we tackle the problem of learning representations for nodes in multiplex networks in an unsupervised or self-supervised manner. To that end, we explore diverse information fusion schemes performed at different levels of the graph processing pipeline. The detailed analysis and experimental evaluation of various scenarios inspired us to propose improvements in how to construct GNN architectures that deal with multiplex graphs.