Explaining Graph Neural Networks with mixed-integer programming

Abstract

Graph Neural Networks (GNNs) provide state-of-the-art graph learning performance, but their lack of transparency hinders our ability to understand and trust them, ultimately limiting the areas where they can be applied. Many methods exist to explain individual predictions made by GNNs, but there are fewer ways to gain more general insight into the patterns they have been trained to identify. Most existing methods for model-level GNN explanations attempt to generate graphs that exemplify these patterns, but the discreteness of graphs and the nonlinearity of deep GNNs make finding such graphs difficult. In this paper, we formulate the search for an explanatory graph as a mixed-integer programming (MIP) problem, in which decision variables specify the explanation graph and the objective function represents the quality of the graph as an explanation for a GNN’s predictions of an entire class in the dataset. This approach, which we call MIPExplainer, allows us to directly optimize over the discrete input space and find globally optimal solutions with a minimal number of hyperparameters. MIPExplainer outperforms existing methods in finding accurate and stable explanations on both synthetic and real-world datasets. Code is available at https://github.com/blake-gaines/MIPExplainer.