PUB - Publikationen an der Universität Bielefeld

Graph Neural Networks (GNNs) have demonstrated state-of-the-art performance in many applications from disease prediction to weather forecasting. However, each GNN layer introduces new trainable parameters that increase its complexity and limit its interpretability. To address these limitations, we propose GraphAware, a new framework that enables an efficient and interpretable analysis of graph-structured data. GraphAware uses easily customizable aggregation functions to aggregate neighborhood features without training which are then passed to standard machine learning classifiers compatible with established interpretability frameworks such as SHAP. We show that GraphAware achieves competitive classification performance compared to state-of-the-art GNNs like Graph Attention Networks on transductive and inductive graph benchmarks. In addition, we demonstrate that the returned results are highly interpretable, improving decision making, transparency, error analysis, and trust in trained models. GraphAware can use popular Python packages such as scikit-learn and XGBoost and uses a scikit-learn-like API to increase usability. The complete framework is open-source on https://github.com/danielwalke/GraphAware.