Introduction
Graph Neural Networks (GNNs) face significant challenges in generalizing across diverse graph structures, specifically homophilic networks (where neighbors are similar) and heterophilic networks (where neighbors differ). While spectral GNNs utilize low-pass and high-pass filters to capture smoothness and differences respectively, existing self-supervised Graph Contrastive Learning (GCL) methods often fail to leverage the inherent opposition between these signals.
LOHA (LOw-pass and High-pAss views) is a novel framework designed to address the "diversity vs. harmony" trade-off. It treats these opposing spectral views as "natural negatives" to maximize feature diversity while maintaining node identity through a novel stabilization mechanism called the Spectral Signal Trend (SST).
The LOHA Methodology
The LOHA framework is built upon three primary pillars designed to maximize feature diversity without sacrificing structural consistency:
- Sliding Filter Construction: Utilizing Chebyshev polynomials, LOHA employs a sliding cosine-parametric formulation. This provides greater flexibility than fixed-interval interpolations, allowing the model to adaptively emphasize critical frequency bands.
- Views Separation (Diversity): In a departure from traditional GCL—which typically pulls different views of the same node together—LOHA treats low-pass and high-pass representations of the same node as negatives. This forces the encoders to extract highly specialized, non-redundant information from each spectral band.
- Views Reunion (Harmony): To prevent feature divergence, the Spectral Signal Trend (SST) is introduced as a "composite feature." Inspired by Dirichlet Energy, SST measures signal variance across a node's local neighborhood, providing a stable anchor that ensures diverse spectral views remain grounded in the same node identity.
Experimental Evaluation and Results
LOHA was rigorously tested across nine real-world datasets, spanning both homophilic (e.g., Cora, PubMed) and heterophilic (e.g., Cornell, Wisconsin) graphs. The model was compared against a comprehensive suite of baselines, including supervised GNNs like ChebNet and BernNet, as well as state-of-the-art GCL frameworks like GRACE and PolyGCL.
Key Findings
- Superior Performance: LOHA achieved an average improvement of 2.8% over the best-performing baselines across all datasets.
- Heterophily Excellence: The model showed remarkable gains on heterophilic graphs, notably improving accuracy on the Wisconsin dataset by 11.9% over the next best self-supervised method.
- Outperforming Supervised Models: In several instances, LOHA's self-supervised approach surpassed fully supervised benchmarks, indicating its effectiveness in capturing structural nuances that label-based training might miss.
Discussion and Conclusion
The success of LOHA underscores the value of "natural negatives" in contrastive learning. By minimizing similarity between opposing spectral views while using the SST as a stabilizing "glue," LOHA achieves a delicate balance between feature diversity and identity consistency. While hyperparameter sensitivity and computational scaling for extremely large graphs remain areas for future work, the "harmony in diversity" principle offers a promising direction for multi-view learning tasks beyond the graph spectral domain.
Summary updated