Algorithm ensures feasible changes in counterfactual explanations
FCx: An algorithm for finding Feasible Counterfactual Explanations
Machine Learning
Summary
Counterfactual explanations show how changing certain features would alter a decision, like getting a loan approval if your income were higher. Existing methods struggle to suggest changes that both make sense and could actually happen. The authors created a new method, FCx, that makes sure the suggested changes are realistic, affordable, and possible by using both expert rules and data-driven insights. Their approach uses a special kind of neural network to find these feasible explanations efficiently. Tests on public datasets show FCx performs well while ensuring feasible suggestions.
What this means in practice
- •For credit risk teams: Generate actionable explanations for loan decisions that suggest realistic and possible changes applicants can make.
- •For human resources analytics: Improve candidate evaluation explanations by ensuring suggested changes respect ethical and legal feasibility constraints.
Authors
Kleopatra Markou, Vana Kalogeraki, Dimitrios Gunopulos
Abstract
Counterfactual (CF) explanations identify changes that alter an input's classification. While existing methods produce realistic and low-cost CFs, they often fail to ensure feasibility, by suggesting non-constructive modifications or incompatible with future changes (e.g., changing an individual's race to secure a job offer). We introduce a refinement of CF explanations that explicitly enforces feasibility. Our approach is the first to efficiently generate CFs that are realistic, low-cost and feasible. We accommodate both hard feasible constraints, specified by domain knowledge users, and soft feasible constraints, inferred automatically via causal inference from the dataset. Our method, Feasible Counterfactual Explanations (FCx), is based on a modified Variational Autoencoder (VAE) optimized with a multi-factor loss function. We measure the cost of a change based on the absolute change in values (proximity) as well as the number of features changed (sparsity) while realism is measured based on the LOF for density estimation, guaranteeing that CFs reside in densely populated regions. Extensive experiments on four public datasets show that our approach matches state-of-the-art performance across multiple metrics while guaranteeing feasibility.