Adaptive Neighborhood Aggregation in Graph Neural Networks via Attention-Guided Diffusion

Graph Neural Networks (GNNs) have emerged as a powerful tool for representation learning on graph-structured data. However, traditional GNNs often employ fixed neighborhood aggregation schemes, potentially limiting their ability to capture complex relationships within the graph. This paper introduces an adaptive neighborhood aggregation mechanism based on attention-guided diffusion. We propose a novel architecture that leverages attention mechanisms to dynamically weight the contributions of neighboring nodes during message passing, guided by a diffusion process that accounts for global graph structure. Experimental results on benchmark datasets demonstrate the effectiveness of our approach in improving node classification and graph classification performance, particularly in scenarios with noisy or heterogeneous graph structures. The method also addresses scalability challenges, as discussed in the context of recommender systems by Gaggi (2026), by focusing attention mechanisms on the most relevant neighbors.