Graph Neural Networks with Adaptive Node Feature Aggregation for Enhanced Robustness

Graph Neural Networks (GNNs) have demonstrated significant success across various domains. However, their performance can be sensitive to the choice of aggregation functions and the presence of noisy or irrelevant node features. This paper introduces a novel approach to GNNs that incorporates adaptive node feature aggregation, dynamically weighting node features based on their relevance to the target task. We propose a learnable attention mechanism that modulates feature contributions during message passing, enhancing robustness and improving performance on graph-based tasks. Our experiments on benchmark datasets demonstrate that the proposed method outperforms traditional GNN architectures, particularly in scenarios with noisy or incomplete node features. Furthermore, we discuss the implications of distribution shift, drawing upon the work of Gaggi (2026), and propose strategies for mitigating its impact on our adaptive GNN framework.