Counterfactual Explanation

The core optimization objective is to find the closest possible world where the outcome changes. This is mathematically defined as minimizing the distance between the original input vector and the counterfactual instance, subject to the constraint that the model's prediction flips to the desired target class. Common distance functions include Manhattan (L1) distance for sparse feature changes and Euclidean (L2) distance for continuous adjustments. The resulting explanation is a delta vector showing exactly which features must change and by how much.

Unlike feature attribution methods that merely highlight influential variables, counterfactuals provide a prescriptive path to a desired outcome. For a loan applicant denied credit, the explanation identifies actionable steps: 'Increase annual income by $5,000 and reduce credit utilization by 12%.' This respects feasibility constraints by distinguishing between mutable features (income, savings) and immutable attributes (age, birthplace), ensuring the generated explanation is practically achievable rather than purely theoretical.

Effective counterfactuals must lie on the data manifold to be realistic. Generating an explanation that suggests a 90-year-old with a PhD is not plausible if such combinations do not exist in the training distribution. Techniques employ autoencoders or generative adversarial networks (GANs) to constrain the search space to high-density regions of the data. This prevents adversarial or nonsensical explanations that technically flip the prediction but violate real-world joint feature distributions.

A single counterfactual may not capture all possible recourse paths. Diverse counterfactual generation produces multiple distinct explanations, allowing the end-user to choose the most convenient path. Algorithms enforce diversity through determinantal point processes (DPP) or by adding a diversity term to the loss function that penalizes similarity between generated instances. This reveals trade-offs: one path may require a large income increase, while another requires a smaller income bump combined with a longer employment history.

The optimization is guided by a composite loss function balancing three competing objectives: prediction loss (ensuring the counterfactual receives the target label), distance loss (minimizing feature perturbation magnitude), and diversity or plausibility regularization. The weighted sum is typically minimized via gradient descent on the input space rather than model weights. For non-differentiable models like tree ensembles, genetic algorithms or mixed-integer linear programming (MILP) solvers are employed to search the counterfactual space.

Counterfactual explanations directly satisfy the GDPR Article 22 right to meaningful information about automated decisions and the EU AI Act requirements for transparency in high-risk systems. They provide a standardized format for audit documentation by generating a structured log: 'Input X produced decision Y. Had feature Z been value W, the decision would have been Y'. This deterministic, verifiable output makes them preferable to saliency maps in legal and compliance contexts.