Using Causal Constraints to Achieve Diverse Feasible and Actionable Counterfactual Explanations Predictions with SENSE-EGRA Dataset

Abstract

This work presents a framework tiled – DiFACE-SCM, which stands for Diverse, Feasible, and Actionable Counterfactual Explanations (CE) with Structural Causal Model constraints. The background knowledge obtained from the data generation process is employed to design a structural causal model (SCM), which is further used to identify variable features which can be used to constrain and sparse a black-box machine learning (ML) predictive model and ensues it produces feasible and actionable counterfactual explanation during the model prediction process, to identify features in the dataset that will require improvement for each student in the next intervention program. The designed SCM is tested for correctness using the conditional independence test (CIT) criteria. Further, it uses a real-life intervention dataset of the American University of Nigeria (AUN) called “Strengthening education in northeast Nigeria (SENSE) early grade reading assessment (EGRA) - (SENSE-EGRA) dataset. A maximum of 4 counterfactual explanations (CEs) are generated for each test dataset instance, and the result shows feasible and actionable CEs for all test dataset instances for the categorical variables features as compared to similar previous frameworks. Albeit, with the numeric variables, the framework sometimes generates CEs that are infeasible and non-actionable. This challenge is however resolved by further introducing non-causal constraints adopted from a previous framework. Also, the constrained sparse ML black-box predictive model led to a reduction in the model’s F1 and accuracy scores. However, the focused of this work is on feasible and actionable CEs and so we cannot tradeoff feasible and actionable CEs on the altar of high accuracy and F1 scores with meaningless CEs.