PolyGCL: GRAPH CONTRASTIVE LEARNING via Learnable Spectral Polynomial Filters

Graph contrastive learning has gained significant attention for learning effective representations from unlabeled graph data. In this paper reproduction, we replicate and evaluate the methodology proposed by the PolyGCL pipeline, which uses spectral polynomial graph filters to achieve contrastive learning by learning a linear combination of a high-pass and low-pass view. The original paper studies the necessity of introducing high-pass information when a graph presents a high degree of heterophily, that is, when nodes connected by an edge are likely to have different labels, which poses a challenge for current unsupervised learning approaches. Through this replication study, we assess the reproducibility and robustness of the method using both their open-source code and our own implementation, discuss implementation details, and provide insights into the practical considerations and challenges of applying self-supervised contrastive learning.