Knowledge and Information Systems Human-Computer Interaction

Structural invariant augmentation for graph contrastive learning based on optimal transport

Xie, Tao, Liu, Cheng, Xu, Zhuopeng, Tan, Jun, Li, Jinglin, Gui, Ning
Springer Nature · 2026-11-06 · DOI 10.1007/s10115-026-02915-x

Abstract

Graph contrastive learning (GCL) learns representations from augmented graph views, whose topology perturbations can distort node-level degree information. We characterize this distortion through marginal structural distributions (MSDs), the normalized row and column sums of a weighted adjacency matrix. We propose OTGCL, which combines Diverse Optimal Transport Augmentation (DOTA) with Marginal Structure-based Sampling (MSS). DOTA constructs topological views through local optimal transport under constraints that preserve each node’s incoming and outgoing weighted degrees. Random subgraph sampling provides diverse views and limits the transport problem size; MSS uses degree information to select within-view negatives. Experiments on eight benchmark graphs show gains in degree prediction and competitive results on node classification and link prediction, together with benefits from using DOTA in the tested GCL backbones. Structural measurements verify weighted-marginal preservation to numerical precision and quantify changes in binary degrees and higher-order properties. Cost and sampling studies reveal graph- and task-dependent responses. Local optimal transport also supports representation learning on larger graphs. Together with computational measurements, these studies characterize the predictive performance and cost of weighted-marginal-preserving augmentation.

Read at publisher →