Convex--Concave Quadratic Spectral Filtering for Graph Neural Networks

Researchers propose DCQ-GNN, a spectral graph neural network using adaptive convex-concave quadratic filters to improve frequency selectivity without high computational costs. The model demonstrates competitive performance on both homophilic and heterophilic graphs while maintaining robustness under structural perturbations.

DCQ-GNN addresses a fundamental trade-off in spectral graph neural networks: low-order filters are computationally efficient but lack spectral selectivity, while high-order polynomial filters offer better selectivity at the cost of optimization complexity. By restricting filter order to two and explicitly leveraging complementary curvature properties, the researchers achieve a middle ground that improves performance without sacrificing stability.

The advancement builds on established spectral filtering theory but applies a novel curvature-aware design philosophy. Rather than simply increasing polynomial degree, DCQ-GNN uses banks of convex and concave quadratic filters that complement each other's frequency responses. This approach reflects a broader trend in machine learning toward more interpretable, theoretically grounded architectural choices that avoid brute-force complexity increases.

For the graph neural network community, this work has practical implications. GNNs are increasingly deployed in recommendation systems, molecular modeling, and network analysis tasks where both computational efficiency and model quality matter. The benchmark results—tying for top performance on heterophilic graphs and ranking second on homophilic ones—suggest DCQ-GNN provides genuine utility rather than marginal improvements. The robustness gains under structural perturbations particularly matter for real-world applications where graph data often contains noise or missing information.

The research direction signals growing sophistication in spectral GNN design beyond simple polynomial expansions. Future work likely explores whether similar curvature-aware principles apply to other graph learning tasks, and whether adaptive gating mechanisms can be further optimized for specific domain structures.

• →DCQ-GNN achieves competitive performance using only quadratic filters, avoiding high-order polynomial optimization challenges
• →Curvature-aware spectral filter design improves selectivity while maintaining computational efficiency
• →The model demonstrates significantly better robustness to structural perturbations than first-order and high-order baselines
• →Node-adaptive gating enables structure-aware spectral filtering tailored to individual graph properties
• →Formal spectral analysis provides theoretical grounding in Dirichlet energy and von Neumann entropy frameworks