Interpretable Graph Intelligence: A Journey from Black to White Box
Mohamed Abdel‐Basset, Nour Moustafa, Hossam Hawash, Zahir Tari · 2023
In several areas of artificial intelligence (AI), deep learning techniques are consistently outperforming their predecessors. The lack of interpretability is one of their main drawbacks of deep learing models. Because of this limitation, the field of interpretability arose to develop post hoc approaches to explain predictions. Interpretability of deep models on images and texts has made great strides recently. Graph neural networks and the methods by which they might be explained are two areas of rapidly expanding research. The field of interpretable graph intelligence, however, lacks a cohesive treatment, as well as a standardized benchmark and testbed for evaluations. This chapter provides taxonomic exploration of the existing approaches to explaining the graph intelligence models. The holistic and taxonomic discussion of interpretability approaches illuminates the similarities and contrasts between current approaches, paving the way for future methodological advancements. The chapter provides some practical implementation of common interpretability methods. By the end, the chapter extends the discussion to cover the metrics for evaluating the explanation of graph networks.