Random forest explainability using counterfactual sets

• Published in: Information fusion 2020-11, Vol.63, p.196-207
• Main Authors: Fernández, Rubén R., Martín de Diego, Isaac, Aceña, Víctor, Fernández-Isabel, Alberto, Moguerza, Javier M.
• Format: Article
• Language: English
• Subjects: Counterfactual; Counterfactual sets; Decision tree; Explainable machine learning; Information fusion; Random forest
• ISSN: 1566-2535; 1872-6305
• DOI: 10.1016/j.inffus.2020.07.001

•Counterfactual sets, a new explanation technique based on counterfactuals.•Counterfactual sets are summarized using an interpretable representation.•A fusion of Random Forest tree predictors into a single Decision Tree.•A method to extract counterfactual sets from a Random Forest.•The extracted counterfactual set contains the optimal counterfactual by design. Nowadays, Machine Learning (ML) models are becoming ubiquitous in today’s society, supporting people with their day-to-day decisions. In this context, Explainable ML is a field of Artificial Intelligence (AI) that focuses on making predictive models and their decisions interpretable by humans, enabling people to trust predictive models and to understand the underlying processes. A counterfactual is an effective type of Explainable ML technique that explains predictions by describing the changes needed in a sample to flip the outcome of the prediction. In this paper, we introduce counterfactual sets, an explanation approach that uses a set of counterfactuals to explain a prediction rather than a single counterfactual, by defining a sub-region of the feature space where the counterfactual holds. A method to extract counterfactual sets from a Random Forest (RF), the RandomForestOptimalCounterfactualSetExtractor(RF−OCSE), is presented. The method is based on a partial fusion of tree predictors from a RF into a single Decision Tree (DT) using a modification of the CART algorithm, and it obtains a counterfactual set that contains the optimal counterfactual. The proposal is validated through several experiments against existing alternatives on ten well-known datasets by comparing the percentage of valid counterfactuals, distance to the factual sample, and counterfactual sets quality.