DisCERN: Discovering Counterfactual Explanations using Relevance Features from Neighbourhoods

Abstract

Counterfactual explanations focus on "actionable knowledge" to help end-users understand how a machine learning outcome could be changed to a more desirable outcome. For this purpose a counterfactual explainer needs to discover input dependencies that relate to outcome changes. Identifying the minimum subset of feature changes needed to action an output change in the decision is an interesting challenge for counterfactual explainers. The DisCERN algorithm introduced in this paper is a case-based counter-factual explainer. Here counterfactuals are formed by replacing feature values from a nearest unlike neighbour (NUN) until an actionable change is observed. We show how widely adopted feature relevance-based explainers (i.e. LIME, SHAP), can inform DisCERN to identify the minimum subset of "actionable features". We demonstrate our DisCERN algorithm on five datasets in a comparative study with the widely used optimisation-based counterfactual approach DiCE. Our results demonstrate that DisCERN is an effective strategy to minimise actionable changes necessary to create good counterfactual explanations.

How Close Is Too Close? The Role of Feature Attributions in Discovering Counterfactual Explanations

“…similar cases with different class labels (see Figure 1) [7,8]. A NUN represents potential changes to the current problem, with feature attribution prioritising the changes that, when actioned, can lead to a different outcome [8,9]. Focusing on a small number of key "actionable" features is more desirable from a practical standpoint, and has the benefit of reducing the recipient's cognitive burden for understanding the counterfactual.…”

Towards Feasible Counterfactual Explanations: A Taxonomy Guided Template-Based NLG Method

“…Sparsity calls for minimising the number of modified features, while proximity ensures that the counterfactual instance is as close as possible to the original instance in the feature space, thereby seeking the minimal change necessary to achieve the desired outcome [12]. Both can be addressed either by case-based instance learning [2,5,27] or as parameters within optimisation minimisation techniques [18,25]. Feasibility ensures suggested changes are achievable [20], and plausibility maintains realistic distributions [28].…”

“…It serves three primary goals [25]: 1) elucidate the reasoning behind decisions; 2) supply adequate information to critique decisions with negative outcomes; and 3) enable a better understanding of the necessary changes to achieve desired outcomes in the future. There is an abundance of techniques to generate cf-XAI in the literature that achieve some subsets of these three goals [2,9,18,24,27]. The focus of this paper instead is to achieve the third goal as a post-processing step taking into account the user perspective.…”

Outcome-Guided Counterfactuals for Reinforcement Learning Agents from a Jointly Trained Generative Latent Space

“…Given a query, if a NUN cannot be found, the closest pair is selected and the featurelevel differences between the examples in that pair are used to transform the query into the target. Another approach uses feature-relevance methods like SHAP (Lundberg and Lee 2017) to tailor a feature edit schedule for converting the query into a counterfactual (Wiratunga et al 2021). In this body of work, an imitating function is trained to predict the outcome variable given the instance features.…”

“…While drawing NUNs can ensure plausibility, there is no guarantee of proximity: There may be no instances that are sufficiently similar to the query. Several studies have generated counterfactuals for tabular data by interpolating between the query and the NUN (Keane and Smyth 2020; Wiratunga et al 2021). As we are using a generative model, we can perform a similar interpolation in the latent space by interpolating linearly between the latent encodings of the query z q and the NUN z N U N to obtain the interpolated latent representation z ι .…”