GeoGRACE: Geometry-Inspired Low-Rank Parameter Perturbation for Efficient Graph Contrastive Learning

Abstract

In molecular property prediction, unsupervised graph contrastive learning (GCL) is pivotal when obtaining labeled data is costly. However, conventional augmentation-free approaches based on parameter perturbation mainly rely on full-rank isotropic noise, which can introduce uncontrolled perturbations and impair semantic consistency of the molecular representations. To mitigate semantic degradation and improve representation consistency, this paper therefore introduces GeoGRACE, a geometry-inspired low-dimensional parameter perturbation framework that generates contrastive views by controlling the amount of noise injected into the parameters in a low-dimensional random subspace using low-rank matrix factorizations. Experiments on five biochemical benchmark datasets show that this structured geometric prior offers improved or comparable downstream classification accuracy over the SimGRACE baseline, for example, improving accuracy from 78.35% to 79.18% on NCI1 dataset. The robustness analysis suggests that early stopping and layer-wise adaptive rank can improve performance stability. Moreover, introducing additional low-rank factorization operations does not necessarily introduce substantial computational overhead; the Layer-wise Adaptive Rank variant, for example, reduces training time by up to 18.5% with manageable memory cost. This work presents a computationally efficient and representation-stable pre-training paradigm that shows promise for biochemical graph representation learning.