Title: Transforming Temporal Snapshots into Structural Models: Graphon-Mixture Approaches for Mixup and Contrastive Learning

Abstract:

Graph datasets encountered in practical applications frequently stem from heterogeneous populations, wherein individual graphs are produced by various distinct underlying distributions. This study introduces a comprehensive framework that treats graph data as a mixture of probabilistic generative models, specifically those defined by graphons. To identify and estimate these graphon structures, we utilize graph moments, such as motif densities, to group together graphs originating from identical underlying models. We present a new theoretical finding that establishes a tighter bound, demonstrating with high probability that graphs drawn from graphons with similar structures will possess comparable motif densities. This insight facilitates the rigorous estimation of the components within graphon mixtures. Furthermore, we illustrate how integrating these estimated mixture components can improve two prominent downstream methodologies: graph contrastive learning and data augmentation through mixup. By anchoring these techniques to their respective generative foundations, we introduce Graphon-Mixture-Aware Mixup (GMAM) and Model-Aware Graph Contrastive Learning (MGCL). Our extensive evaluations on both synthetic and real-world data reveal robust empirical results. In the context of supervised learning, GMAM surpasses current augmentation methods, setting new state-of-the-art accuracy records on six out of seven tested datasets. For unsupervised learning, MGCL demonstrates competitive performance across seven benchmark datasets, securing the lowest overall average rank.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC