Search papers, labs, and topics across Lattice.
This paper introduces a two-view framework for node prediction that aligns structure-conditioned Graph Neural Network (GNN) embeddings with a feature prior from an independent anchor model, addressing vulnerabilities in conventional GNNs caused by topology noise and heterophilous connections. The authors propose a Channel-Split Adaptive Gated GNN (CSAG-GNN) that employs a node-wise gating mechanism to dynamically balance global spectral smoothing and local spatial discrimination. Empirical evaluations demonstrate that their approach achieves significant performance improvements and enhanced structural robustness across various benchmarks compared to existing methods.
GNNs can now maintain structural robustness while effectively handling noise and heterophily, thanks to a novel two-view framework that decouples feature transformation from neighborhood aggregation.
We introduce a constrained two-view framework for node prediction that aligns structure-conditioned GNN embeddings with a structure-free feature prior learned by an anchor model. Conventional Graph Neural Networks (GNNs) couple feature transformation and neighborhood aggregation, which renders them vulnerable to topology noise and heterophilous connections. To decouple this dependency, our framework utilizes an independent anchor network to capture intrinsic attribute features via a self-supervised reconstruction objective. Furthermore, we propose a Channel-Split Adaptive Gated GNN (CSAG-GNN) that dynamically routes representations between global spectral smoothing and local spatial discrimination through a node-wise gating mechanism. We propose a stable cyclic alternating optimization strategy to solve the resulting coupled bi-level objective, preventing mutual representation drift during training. Empirical results on both homophilous and heterophilous benchmarks show balanced performance gains and structural robustness over competitive baselines.
