Consistent Counterfactuals for Deep Models

Abstract

Counterfactual examples are one of the most commonly-cited methods for explaining the predictions of machine learning models in key areas such as finance and medical diagnosis. Counterfactuals are often discussed under the assumption that the model on which they will be used is static, but in deployment models may be periodically retrained or fine-tuned. This paper studies the consistency of model prediction on counterfactual examples in deep networks under small changes to initial training conditions, such as weight initialization and leave-one-out variations in data, as often occurs during model deployment. We demonstrate experimentally that counterfactual examples for deep models are often inconsistent across such small changes, and that increasing the cost of the counterfactual, a stability-enhancing mitigation suggested by prior work in the context of simpler models, is not a reliable heuristic in deep networks. Rather, our analysis shows that a model's Lipschitz continuity around the counterfactual, along with confidence of its prediction, is key to its consistency across related models. To this end, we propose Stable Neighbor Search as a way to generate more consistent counterfactual explanations, and illustrate the effectiveness of this approach on several benchmark datasets.

Robust Counterfactual Explanations for Tree-Based Ensembles

“…However, this work focuses on small changes to the model, e.g., retraining on some data drawn from the same distribution, or minor changes to the hyperparameters, keeping the underlying data mostly similar. Such small changes to the model are in fact quite common in several applications and occur frequently in practice [12][13][14][15].…”

“…In [6,10,18], the authors argue that counterfactuals that lie on the data manifold are likely to be more robust than the closest counterfactuals, but the focus is more on generating counterfactuals that specifically lie on the data manifold (which may not always be sufficient for robustness). Despite researchers arguing that robustness is an important desideratum of local explanation methods [13], the problem of generating robust counterfactuals has been less explored, with the notable exceptions of some recent works [12,14,22]. In [12,14], the authors propose algorithms that aim to find the closest counterfactuals that are also robust (with demonstration on linear models and neural networks).…”

“…Despite researchers arguing that robustness is an important desideratum of local explanation methods [13], the problem of generating robust counterfactuals has been less explored, with the notable exceptions of some recent works [12,14,22]. In [12,14], the authors propose algorithms that aim to find the closest counterfactuals that are also robust (with demonstration on linear models and neural networks). In [22], the focus is on analytical trade-offs between validity and cost.…”

Algorithmic Recourse in the Face of Noisy Human Responses

“…Prior works have focused on determining the extent to which recourses remain robust to the choice of the underlying model (Pawelczyk et al, 2020b;Black et al, 2021), shifts or changes in the underlying models (Rawal et al, 2021;Upadhyay et al, 2021), or small perturbations to the input instances (Artelt et al, 2021;Dominguez-Olmedo et al, 2021;Slack et al, 2021).…”

On the Adversarial Robustness of Causal Algorithmic Recourse

“…lead to favourable classification outcomes) for all plausible individuals similar to the individual seeking recourse. We refer to this notion of robustness as the adversarial robustness of recourse, in order to distinguish it from other robustness considerations previously studied in the recourse literature (e.g., robustness with respect to changes to the decision-making classifier [29,36,3]), and as a reference to the adversarial robustness literature, which considers robustness of prediction precisely against uncertainty in the features of the data.…”