Optimal Recourse Summaries via Bi-Objective Decision Tree Learning

Abstract: Actionable Recourse provides individuals with actions they can take to change an unfavorable classifier outcome. While useful at the instance level, it is ill-suited for global auditing and bias detection, since aggregating local actions is costly and often inconsistent. Recourse Summaries address this limitation by partitioning the population and assigning one shared action per subgroup, enabling comparison across subgroups. Designing summaries involves a fundamental trade-off between recourse effectiveness and recourse cost, which existing methods do not adequately address. We introduce Summaries of Optimal and Global Actionable Recourse (SOGAR), which formulates recourse summary learning as an optimal decision tree learning problem and finds the Pareto front -- the complete set of solutions where improving one objective necessarily worsens the other. SOGAR enables post-hoc selection of the desired trade-off without retraining. Using shallow axis-parallel decision trees and sparse leaf actions, SOGAR produces stable, low-cost, and effective recourse summaries that outperform existing approaches across effectiveness and cost metrics.

• The paper presents the SOGAR framework, which computes globally optimal recourse summaries by simultaneously minimizing recourse cost and loss.
• It leverages STreeD’s dynamic programming and caching strategies to efficiently enumerate a full Pareto front of non-dominated, interpretable solutions.
• Empirical evaluations show SOGAR significantly reduces invalid recourse and exposes bias, offering robust auditing for fairness in decision-making.

Optimal Recourse Summaries via Bi-Objective Decision Tree Learning: A Technical Essay

Motivation and Problem Formulation

The paper "Optimal Recourse Summaries via Bi-Objective Decision Tree Learning" (2605.07598) addresses a key limitation in algorithmic recourse: the gap between individualized, instance-level explanations and the need for global, population-level auditing to identify patterns such as bias or systemic inefficiency. Existing individualized recourse methods provide actionable feedback for specific subjects but lack coherence when aggregated for subgroup or global analysis, hindering practical auditing and bias detection. Recourse summaries—assigning a shared action to each subgroup—attempt to bridge this gap but, to date, have not achieved globally optimal trade-offs between recourse loss (ineffectiveness) and recourse cost, instead often relying on arbitrary scalarizations of the objectives or heuristics with no global optimality guarantees.

The SOGAR Framework

The proposed SOGAR (Summaries of Optimal and Global Actionable Recourse) framework conceptualizes recourse summary learning as a bi-objective, globally optimal decision tree problem, leveraging the recent STreeD dynamic programming approach for optimal tree learning. Specifically:

• Recourse Summaries as Decision Trees: SOGAR constructs shallow, axis-parallel decision trees whose leaves define interpretable subgroups; each leaf is assigned a single, sparse action that is feasible for that subgroup. This tree structure controls subgroup count and description complexity.
• Bi-objective Optimization: SOGAR simultaneously minimizes recourse cost—measured using the Maximum Percentile Shift (MPS) for scale invariance—and recourse loss—the frequency of assigned actions failing to flip outcomes. Rather than commit to a single cost-effectiveness trade-off, SOGAR computes the full Pareto front of non-dominated solutions, allowing downstream users to select from the spectrum of possible trade-offs without retraining.
• Optimization Guarantees: The problem is cast within the separable task class of STreeD, enabling globally optimal Pareto front enumeration up to practical problem size limitations.

Algorithmic and Computational Advances

Key technical contributions include:

• Action & Cost-Loss Caching: SOGAR precomputes and caches the cost and loss for every action-instance pair, amortizing repeated computations during tree search. This is critical for practical scalability given the combinatorial search space.
• Dynamic Programming via STreeD: STreeD's bottom-up DP with state caching and strong pruning drastically reduces the number of subproblems, enabling exploration of non-greedy, globally-optimal trees with manageable depth and leaf constraints.
• Parallelization: The evaluation of action assignments per leaf is embarrassingly parallel; the paper demonstrates significant wall-clock time reductions using multi-thread CPU and GPU acceleration.
• Anytime Solution: SOGAR enables early termination after a time budget, returning the best non-dominated summaries found so far—key for scalability to large datasets.

Experimental Evaluation

The empirical evaluation comprises four tabular datasets relevant for recourse and fairness auditing: Employee Attrition, German Credit, Bank Marketing, and Adult Income, using state-of-the-art classifiers (LightGBM, XGBoost, DNN). Each dataset is processed with immutability constraints on sensitive features, constraining actions to realistic interventions.

Baselines include: AReS, CET, GLOBE-CE, GLANCE, and T-CREx. Comparisons utilize cross-validation and standard recourse metrics: cost (MPS), loss, invalidity (cost+loss), and runtime.

Strong numerical results include:

• Invalidity minimization: SOGAR delivers the lowest average invalidity across all datasets, often by a substantial margin, and consistently beats scalarization-based and heuristic baselines on this metric.
• Pareto front density and flexibility: SOGAR yields thousands of non-dominated trade-offs in a single run (e.g., ~6,000 solutions for Adult), versus a single solution from most baselines. This provides greater flexibility for auditors and policy-makers.
• Auditing performance: SOGAR robustly recovers known biases; in the Adult dataset, it exposes persistent gender disparities in recourse cost and loss across the entire Pareto front—a diagnostic that single-solution methods often miss.
• Overfitting avoidance: Pareto front solutions generalize well to held-out data, with near-invariance in cost/loss trade-offs, indicating robustness of the underlying optimization.

Efficiency trade-off: Although SOGAR incurs higher total computation time than cluster-based or vector-translation methods (GLANCE, GLOBE-CE), its wall-time per solution is competitive (e.g., ~0.2s/solution for Adult), and it eliminates repeated retraining for different trade-offs. Additionally, its anytime property and depth/action sparsity/bins ablations enable practical adaptation under compute constraints.

Theoretical Properties and Optimality

A foundational contribution is the proof that recourse summary tree generation is a separable task suitable for globally optimal dynamic programming [32]. The cost and loss objectives are aggregable and branch-local, and combining solutions preserves dominance order, ensuring that the full Pareto front can be constructed without resorting to approximate or greedy search policies.

Notably, SOGAR is the first method that guarantees global optimality of recourse summaries in both cost and loss for constructed subgroups, unlike prior approaches limited by local search, scalarization, or heuristic pruning.

Practical and Theoretical Implications

For model auditing and fairness: SOGAR’s output—interpretable, globally-optimal recourse summaries and their full trade-off frontier—directly supports both regulatory compliance (e.g., Article 86 of the EU AI Act) and practical auditing. The explicit subgroup structure and per-group action assignments facilitate bias detection, group comparability, and actionable policy recommendations. The persistent group disparities revealed in the Pareto front (e.g., for gender in the Adult dataset) illustrate both the value and necessity of multi-objective recourse analysis for fairness.

For algorithmic recourse research: The SOGAR approach demonstrates that globally optimal, interpretable recourse summaries are attainable at moderate tree depth and sparsity constraints, challenging the dominance of heuristic or black-box recourse-generation pipelines. This provides a benchmark for future recourse/summary methods in terms of optimality and interpretability.

Limitations and Future Directions

Despite its advantages, SOGAR has several limitations:

• Computational scalability: While practical for moderately large tabular datasets, the method’s complexity in the number of features, the action space, and the choice of maximum depth/leaves still restricts applicability to very high-dimensional or massive datasets without strong pre-processing or further algorithmic advances.
• Cache memory: The action-instance cache, central to runtime efficiency, may become prohibitive for dense action spaces or large $|D_0|$.
• Dimension of fairness: SOGAR supports fairness auditing by subgroup cost/loss analysis, but fairness remains a multi-dimensional notion; intersectional and indirect forms of bias may require composite or multi-stage auditing.
• Interpretability versus expressivity: Restricting tree depth and action sparsity favors interpretability and computational feasibility, but may limit the ability to capture complex or high-order interactions in the feature space.

Potential extensions include scalable relaxations for ultra-large datasets, structured or hierarchical subgrouping for deeper auditing, and integration with causal constraints for improved action feasibility or regulatory compliance.

Conclusion

SOGAR constitutes a significant advance in recourse summary methodology, providing the first practical framework for globally optimal, interpretable bi-objective recourse summarization via decision trees. Its construction of a full Pareto frontier, together with strong empirical and theoretical properties, directly supports advanced auditing, fairness analysis, and actionable insight in automated decision-making. Future research should focus on scaling such approaches to ever-larger, higher-dimensional, and more complex domains while maintaining guarantees of optimality and transparency.