Graph Neural Network Attribution

Graph neural network (GNN) attribution focuses on understanding how GNNs arrive at their predictions, aiming to improve interpretability and trust in these powerful models. Current research emphasizes developing methods to accurately assign importance scores to individual nodes, edges, or features within the input graph, often employing techniques like gradient-based methods or knowledge distillation within multi-level frameworks. This work is crucial for building reliable GNN-based systems across diverse applications, from high-energy physics to node classification, by enhancing model transparency and facilitating the identification of influential data points or structural patterns. Challenges remain in ensuring the robustness and consistency of attribution methods across different GNN architectures and datasets.