DiCE (Diverse Counterfactual Explanations)

DiCE is implemented as an open-source Python library by Microsoft Research. It provides a high-level API that wraps the optimization logic. Users can generate counterfactuals with a few lines of code by specifying the model, query instance, desired outcome class, and the number of counterfactuals. The library supports both classification and regression tasks and includes built-in support for standard datasets like Adult Income and German Credit.

The foundational concept upon which DiCE is built. A counterfactual explanation describes a causal scenario: 'If input X had been different, prediction Y would not have occurred.' It identifies the minimal set of changes to an instance's features required to flip the model's decision. Unlike feature attribution methods, counterfactuals provide actionable recourse by telling a user exactly what to change. Key properties include validity (the new instance gets the desired outcome), proximity (changes are minimal), and sparsity (few features are changed).

A game-theoretic framework that assigns each feature an importance value—a Shapley value—for a specific prediction. SHAP unifies several additive feature attribution methods and satisfies three desirable properties: local accuracy, missingness, and consistency. While SHAP explains why a prediction was made by decomposing it into feature contributions, DiCE explains what to change to get a different outcome. They are complementary: SHAP for understanding, DiCE for recourse.

A technique that explains individual predictions by approximating the black-box model locally with an interpretable surrogate, such as a sparse linear model. LIME generates perturbed samples around the instance of interest, weights them by proximity, and fits a simple model. The resulting feature weights provide a local explanation. Unlike DiCE, LIME does not natively generate diverse alternative scenarios; it focuses on explaining the current decision boundary rather than exploring paths across it.

A model-agnostic method that provides high-precision rules called anchors. An anchor is a set of feature conditions that sufficiently 'anchors' a prediction locally—meaning changes to other features do not alter the outcome. For example: 'If age > 60 and income < $30k, the loan is denied regardless of other factors.' While DiCE generates actionable instances, Anchors define the decision boundary region where the prediction is invariant, helping users understand the scope of necessary changes.

An adaptation of standard Shapley values that incorporates a Structural Causal Model (SCM) of the data-generating process. Standard SHAP values measure statistical association, which can be misleading when features are correlated. Causal Shapley values use interventional distributions to break spurious correlations, assigning importance based on true causal effects. This connects directly to DiCE's requirement for feasible counterfactuals—changes must respect causal constraints to be actionable in the real world.