Learning Relational Invariance for Out-of-Distribution Molecular Relational Learning
Keywords: molecular relational learning, out of distribution, invariant learning
Abstract: Molecular Relational Learning (MRL) expands the scope of molecular representation learning by incorporating additional molecules, aiming to understand the interactions between pairs of molecules. While MRL has shown promising results, the existing methods have not been able to generalise to real world scenarios. Invariant learning is pivotal in addressing Out-of-Distribution (OOD) generalization challenges. However, two major obstacles impede the progress of invariant learning in MRL: (1) Unlike single-molecular cases, interactions between molecules introduce added complexity, with a heavy reliance on molecular substructure recognition, often leading to the misspecification of invariant patterns. (2) Accurate modeling of interactions can effectively improve generalizations. However, previous methods focus on node interaction, which is limited by the expressiveness of GNN, and long-range interactions cannot be captured. To address these, we propose a novel Relational Invariant Learning (RIL) framework that uses a multi-granularity interaction approach to improve OOD generalization for MRL, and the framework is denoted as RILOOD. Specifically, we model the environment diversity distribution of molecules by mixup-based Conditional Modeling. Then, we employ a multi-granularity refinement strategy to learn the Context-Aware Representation, which is essential for capturing multi-level interaction. We further design an invariant learning module to capture the invariant patterns that robustly generalize across unseen environments. Extensive experiments on molecular datasets show that our method achieves stronger generalization against state-of-the-art methods in the presence of various distribution shifts. Our code will be released after our paper is accepted.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 614
Loading