Counterfactual Explanations

Counterfactual explanations ask what minimal or feasible changes to an instance would make a model produce a different prediction, helping users inspect decisions and identify possible recourse. Methods range from gradient or evolutionary optimization and nearest-neighbor search to formal solvers, causal models, and generative approaches such as VAEs, diffusion, and flows; many impose constraints on immutable features, costs, temporal order, or domain plausibility. Work spans tabular decisions as well as images, graphs, time series, recommenders, clinical systems, and autonomous agents, with evaluation balancing validity, sparsity, proximity, plausibility, diversity, and robustness. A central tension is that a valid model flip may be unrealistic or infeasible for a person; robustness depends on which model changes are tested, and automatic quality metrics do not always match human judgments.