Towards Xai for the Facilitation of Human-Machine Learning of Graph Structure
Julia Handl, Nikolay Mehandjiev · 2025
Graph neural networks (GNNs) provide an elegant tool for the formulation of machine learning problems involving relational data. Standard GNN benchmarks frequently assume that an underlying graph structure is readily accessible, e.g., in the form of a traffic or social network. Yet, there is increasing interest in the potential of GNNs in application domains where the graph itself is not pre-defined, or may need to be refined, as part of the learning process. Largely, such work on graph structure learning focuses on data-driven, automated approaches, i.e. methodologies that derive plausible relations using heuristics that draw on feature or label information, or optimize the final GNN's generalization performance. Without the introduction of problem-specific insight, the task of learning the optimal graph structure for a given GNN model is intractable. Furthermore, current understanding of the interaction between graph structure and model performance is poor. We conjecture that the design of explainability tools for elucidating the impact of a GNN's graph structure is important for two reasons: (i) to foster the development of new, robust heuristics for effective graph construction; and (ii) to facilitate the elicitation and integration of application-specific domain knowledge, from human experts, during the process of graph construction. Here, we provide an introduction to, and discussion of, this challenge, with illustrative results on the established Cora benchmark.