ZORRO: Valid, Sparse, and Stable Explanations in Graph Neural Networks

With the ever-increasing popularity and applications of graph neural networks, several proposals have been made to explain understand decisions a network. Explanations for networks differ in principle from other input settings. It is important attribute decision features related instances connected by structure. We find that previous explanation generation approaches maximize mutual information between label distribution produced model be restrictive. Specifically, existing do not enforce explanations valid, sparse, or robust perturbations. In this paper, we lay down some fundamental principles an method should follow introduce metric RDT-Fidelity as measure explanation's effectiveness. propose novel approach Zorro based on xmlns:xlink="http://www.w3.org/1999/xlink">rate-distortion theory uses simple combinatorial procedure optimize RDT-Fidelity. Extensive experiments real synthetic datasets reveal produces sparser, stable, more faithful than network approaches.