Multiplex Graph Contrastive Learning with Soft Negatives

TL;DR

This work tackles cross-scale graph contrastive learning by introducing MUX-GCL, which leverages multiplex representations as effective patches and a patch-affinity mechanism to minimize information loss from false negatives. The framework combines Multiplex Patch Contrast (MPC) with Patch Affinity Estimation (PAE), weighting negative pairs by topological affinities derived from Node2Vec or VGAE, and provides a theoretical lower bound showing the objective tightens mutual information lower bounds beyond prior methods like GRACE. Empirically, MUX-GCL achieves state-of-the-art results on node classification and clustering across five public datasets, with favorable runtime characteristics. The approach offers a principled, scalable path to preserving cross-scale, consistent information in graph representations and delivers code for reproducibility.

Abstract

Graph Contrastive Learning (GCL) seeks to learn nodal or graph representations that contain maximal consistent information from graph-structured data. While node-level contrasting modes are dominating, some efforts commence to explore consistency across different scales. Yet, they tend to lose consistent information and be contaminated by disturbing features. Here, we introduce MUX-GCL, a novel cross-scale contrastive learning paradigm that utilizes multiplex representations as effective patches. While this learning mode minimizes contaminating noises, a commensurate contrasting strategy using positional affinities further avoids information loss by correcting false negative pairs across scales. Extensive downstream experiments demonstrate that MUX-GCL yields multiple state-of-the-art results on public datasets. Our theoretical analysis further guarantees the new objective function as a stricter lower bound of mutual information of raw input features and output embeddings, which rationalizes this paradigm. Code is available at https://github.com/MUX-GCL/Code.

Figures (3)

• Figure 1: Similarity distributions of cross-layer embeddings between two augmented views (for GRACE). All positive pairs are substantially more similar than negative pairs, labeled as $u_mv_n~pos/neg$ with $m$ and $n$ being the layers.
• Figure 2: Overall architecture of MUX-GCL. Contrasts are executed between "effective patches" constructed from all representations of the multiplex encoder, as illustrated by the links. The pairwise affinities of topological embedding estimate the likelihood of being false negatives. Augmentations are implemented as in GRACE.
• Figure 3: Distributions of $T_{D,ij}^{Lk}$ (left) and $T_{S,ij}^{Lk}$ (right) for Cora fitted by Gaussian curves. Results are shown for epoch 300, but are consistent for the entire training process.