ASPECT: Node-Level Adaptive Spectral Fusion for Graph Contrastive Learning

Researchers introduce ASPECT, a novel spectral graph contrastive learning method that adaptively fuses low- and high-frequency graph signals at the node level rather than uniformly across entire graphs. The approach demonstrates improved representation quality on both homophilic and heterophilic graph benchmarks, addressing limitations in existing graph-level fusion strategies.

ASPECT addresses a fundamental limitation in spectral graph contrastive learning by recognizing that different nodes within a graph may benefit from different spectral characteristics. Traditional graph-level fusion approaches apply uniform mixing of low- and high-frequency signals across all nodes, which proves suboptimal when nodes have heterogeneous spectral preferences. This research demonstrates that such uniform strategies incur irreducible regret on mixed-spectral graphs, establishing theoretical justification for node-adaptive approaches.

The method works by learning node-wise spectral policies that determine how each node combines frequency components, then regularizes these policies using channel-wise contrastive evidence. This enables fine-grained adaptation where heterophilic nodes (those with dissimilar neighbors) and homophilic nodes (those with similar neighbors) can independently optimize their spectral mixtures. The extension ASPECT-S adds stability awareness through generated perturbations and Rayleigh-based spectral search bias, improving robustness against graph structure and feature variations.

This advancement impacts graph neural network development and applications in recommendation systems, social networks, and knowledge graphs where spectral properties vary significantly across node populations. The research contributes to improving GNN representation quality without requiring architectural modifications, making it compatible with existing frameworks. The demonstrated improvements on heterophilic benchmarks are particularly valuable since these graphs present greater challenges for traditional GNN methods.

• →ASPECT enables node-level adaptive fusion of graph spectral components instead of applying uniform graph-level policies
• →The method demonstrates improved representation quality on both homophilic and heterophilic graph benchmarks
• →Channel-wise contrastive regularization allows different nodes to learn optimal spectral preferences autonomously
• →ASPECT-S extension improves stability and robustness through generated perturbations and spectral sensitivity analysis
• →The approach maintains theoretical grounding by addressing irreducible regret problems in mixed-spectral graphs