Counterfactual Graphs for Explainable Classification of Brain Networks

Abstract

Training graph classifiers able to distinguish between healthy brains and dysfunctional ones, can help identifying substructures associated to specific cognitive phenotypes. However, the mere predictive power of the graph classifier is of limited interest to the neuroscientists, which have plenty of tools for the diagnosis of specific mental disorders. What matters is the interpretation of the model, as it can provide novel insights and new hypotheses.In this paper we propose counterfactual graphs as a way to produce local post-hoc explanations of any black-box graph classifier. Given a graph and a black-box, a counterfactual is a graph which, while having high structural similarity with the original graph, is classified by the black-box in a different class. We propose and empirically compare several strategies for counterfactual graph search. Our experiments against a white-box classifier with known optimal counterfactual, show that our methods, although heuristic, can produce counterfactuals very close to the optimal one. Finally, we show how to use counterfactual graphs to build global explanations correctly capturing the behaviour of different black-box classifiers and providing interesting insights for the neuroscientists.

Trustworthy Graph Neural Networks: Aspects, Methods and Trends

“…For example, due to the characteristics of graph data, traditional explanation methods for deep learning (e.g., input optimisation methods [137] and soft mask learning methods [138]) cannot be directly applied to GNNs because of the irregularity and discrete topology of graphs. When conducting evaluations, domain knowledge (e.g., brain connectomics [127]) is sometimes necessary to validate GNN explanations.…”

“…Consequently, building trustworthy GNNs requires insights into why GNNs make particular predictions, which has driven an increase in research into the interpretability and explainability of GNNs. These abilities enable researchers to capture causality in GNNs [126] or insights for further investigation in applications [127], foster the implementation of robust GNN systems by developers [128], and guide regulators to ensure the fairness of GNNs [129].…”

“…The group-level methods (e.g., GNNExplainer [22]) generate or choose a prototype sample as the explanation for a group of graph samples. The class-level methods use graph patterns (e.g., XGNN [132]) or statistics (e.g., OFS+OBS [127]) to explain the outputs of GNNs for specified classes. Compared with group-level and class-level methods, the explanations provided by instance-level methods are both more fine-grained and more precise since explanations for a single sample are tailored to explain the sample in question.…”

“…In DEGREE, based on the forward operation decomposition on GNN layers, the node contribution is calculated as the relative contribution (like the Shapely value [130]) when adding the target node to its contexts. Abrate and Bonchi [127] propose a heuristic-search method called OFS+OBS that finds counterfactual graphs to explain graph classification results on brain networks. Lin et al [155] propose a framework called Gem to train an explanation generator; under this approach, the ground-truth explanations in training data are distilled through combining predefined rules and queries on target GNNs.…”

GRAPHSHAP: Motif-based Explanations for Black-box Graph Classifiers

“…In this paper, instead, we consider a specific type of graph classification that arises in many application domains, in which a specific node id corresponds to the same entity in all the input networks: for instance, in classification of brain networks, the same node id represents the same brain region in all the input graphs. We call this setting graph classification with node identity awareness [15,20,1]. While not every application domain requires node identity awareness, it is crucial to exploit these property whenever it occurs, as ignoring it represents an important loss of information.…”