Benchmarking and Rethinking Multiplex Graphs
Keywords: graph neural networks, multiplex graphs, graph benchmark, datasets
Abstract: Multiplex graphs, which represent complex real-world relationships, have recently garnered significant research interest. However, contemporary methods exhibit variations in implementations and settings, lacking a unified benchmark for fair comparison. Additionally, existing multiplex graph datasets suffer from small-scale issues and a lack of representative features. Furthermore, current evaluation metrics are restricted to node classification and clustering tasks, lacking evaluations on edge-level tasks. These obstacles impede the further development of the multiplex graph learning community. To address these issues, we first conducted a fair comparison based on existing settings, finding that current methods are approaching performance saturation on existing datasets with minimal differences; and simple end-to-end models sometimes achieve better results. Subsequently, we proposed a unified multiplex graph benchmark called MGB. MGB includes ten baseline models with unified implementations, formalizes seven existing datasets, introduces four new datasets with text attributes, and proposes two novel edge-level evaluation tasks. Experiments on MGB revealed that the performance of existing methods significantly diminishes on new challenging datasets and tasks. Additional results suggest that models with global attention and stronger expressive power in end-to-end solutions hold promise for future work. The data, code, and documentations are publicly available at https://anonymous.4open.science/r/multiplex-F150.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8718
Loading