Diverse Counterfactuals

A probabilistic mechanism used to enforce diversity in the generated set. DPPs model the probability of selecting a subset of counterfactuals, assigning higher probability to subsets where items are dissimilar. This mathematically penalizes redundancy, ensuring the algorithm explores distinct regions of the feature space rather than returning minor variations of the same explanation.

While standard counterfactuals optimize for proximity (minimal change), diverse counterfactuals prioritize coverage. The goal shifts from finding the single closest point on the decision boundary to mapping the boundary's topology. This provides a user with a menu of trade-offs, such as:

• A high-cost, immediate change
• A low-cost, gradual change
• A change affecting feature A vs. feature B

The selection of a diverse set is often framed as a submodular maximization problem. A utility function measures the value of a set of counterfactuals, balancing relevance (validity and proximity) with diversity. Because submodular functions exhibit diminishing returns, a greedy algorithm can efficiently construct a near-optimal diverse set by iteratively adding the counterfactual that provides the maximum marginal gain in utility.

For high-dimensional data like images, diversity is often enforced in a learned latent space (e.g., using a Variational Autoencoder). Instead of perturbing raw pixels, the algorithm searches for distinct latent codes that decode into valid counterfactuals. This ensures that the diverse explanations are semantically meaningful and lie on the data manifold, avoiding unrealistic adversarial examples that merely exploit pixel-level noise.

Diversity can be explicitly structured around an action set. The algorithm partitions the feasible actions into distinct clusters (e.g., 'education-based changes' vs. 'financial-based changes') and generates at least one counterfactual per cluster. This guarantees that the user receives a portfolio of options spanning different life domains, directly addressing the feasibility of recourse rather than just mathematical diversity.

The Maximum Mean Discrepancy (MMD) can be used as a statistical constraint to enforce diversity. The algorithm minimizes the MMD between the generated set of counterfactuals and a target distribution that represents ideal diversity. This prevents the set from collapsing into a single mode of the data distribution and ensures the explanations are representative of the full range of possible recourse paths available to the user.

A loan applicant rejected for a high debt-to-income ratio receives multiple distinct paths to approval:

• Path A: Reduce revolving credit card balance by $4,200.
• Path B: Increase annual gross income by $8,500.
• Path C: Close two unused credit lines and reduce balance by $1,800.

This prevents a single infeasible recommendation (e.g., 'increase income by $20k') and lets the applicant choose the most actionable recourse based on their circumstances.

A graduate school applicant receives a rejection prediction. A determinantal point process (DPP) generates diverse counterfactuals:

• Path A: Increase GRE quantitative score from 158 to 165.
• Path B: Add one first-author publication while maintaining current scores.
• Path C: Gain 2 years of relevant research experience.

Each path represents a distinct subspace of the feature space, ensuring the student can pursue the option aligned with their timeline and resources.

A clinical decision support model predicts high risk of diabetes complications. Diverse counterfactuals offer alternative intervention strategies:

• Path A: Reduce HbA1c from 8.2% to 6.8% through medication adjustment.
• Path B: Lower BMI from 32 to 28 through lifestyle intervention.
• Path C: Combination of moderate HbA1c reduction (to 7.3%) and blood pressure normalization.

Physicians evaluate feasibility constraints like patient adherence and contraindications before recommending a specific path.

A recidivism risk model flags an individual as high-risk. Diverse counterfactuals identify distinct rehabilitation pathways:

• Path A: Complete substance abuse treatment program and secure stable housing.
• Path B: Obtain full-time employment for 12 consecutive months.
• Path C: Complete GED and enroll in vocational training.

Diversity ensures recommendations don't inadvertently encode algorithmic bias by forcing all individuals toward a single socioeconomic path that may be structurally inaccessible.

A consumer with a 620 FICO score seeks mortgage qualification at 680+. Diverse counterfactuals generated via gradient-based optimization with diversity constraints:

• Path A: Reduce credit utilization from 78% to 30%.
• Path B: Remove two collection accounts and maintain on-time payments for 6 months.
• Path C: Increase average account age by 18 months (no new credit inquiries).

The action set is constrained to only modifiable features, excluding immutable attributes like credit history length.

A resume screening model rejects a candidate. Diverse counterfactuals reveal multiple skill acquisition paths:

• Path A: Obtain AWS Solutions Architect certification.
• Path B: Complete 3 open-source contributions to relevant GitHub repositories.
• Path C: Gain 1 year of experience with Kubernetes in current role.

Diversity is critical here to avoid proxy discrimination—ensuring the model doesn't systematically recommend paths correlated with protected attributes.