Long-range Meta-path Search through Progressive Sampling on Large-scale Heterogeneous Information Networks

Chao Li, Zijie Guo, Qiuting He, Hao Xu, Kun He

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Keywords: Neural architecture search, heterogeneous graph neural networks, long-range dependency, meta-path search

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We investigate the importance of different meta-paths and propose an automatic framework for utilizing long-range dependency in heterogeneous information networks, called Long-range Meta-path Search through Progressive Sampling (LMSPS).

Abstract: Utilizing long-range dependency, though extensively studied in homogeneous graphs, is rarely studied in large-scale heterogeneous information networks (HINs), whose main challenge is the high costs and the difficulty in utilizing effective information. To this end, we investigate the importance of different meta-paths and propose an automatic framework for utilizing long-range dependency in HINs, called Long-range Meta-path Search through Progressive Sampling (LMSPS). Specifically, to discover meta-paths for various datasets or tasks without prior, we develop a search space with all target-node-related meta-paths. With a progressive sampling algorithm, we dynamically shrink the search space with hop-independent time complexity, leading to a compact search space driven by the current HIN and task. Utilizing a sampling evaluation strategy as the guidance, we conduct a specialized and expressive meta-path selection. Extensive experiments on eight heterogeneous datasets demonstrate that LMSPS discovers effective long-range meta-paths and outperforms state-of-the-art models. Besides, it ranks top-1 on the leaderboards of ogbn-mag in Open Graph Benchmark.

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: 3453