PaperlayerView exact revision on arXiv

Source-linked AI summary

Actionable CBFI: Integrating Structural Decomposition and Causal Counterfactual Recourse for Tabular Machine Learning

Sejong Oh

arXiv:2608.27821v1cs.LGcs.AI

TL;DR

Existing counterfactual recourse methods either search broadly over causal interventions or rely on additive attributions that can miss nonlinear feature synergies. A-CBFI integrates localized structural diagnosis with SCM-based prescription, separating active interventions from downstream effects and guiding search toward causal bottlenecks. Across financial and healthcare evaluations, it reduced active intervention burden while preserving recourse cost and convergence on causally feasible instances.

  • Problem

    Existing recourse approaches face diffuse intervention search and additive target selection that may miss higher-order synergies in nonlinear models.

  • Method

    A-CBFI diagnoses synergistic bottlenecks using localized structural decomposition, separates active interventions from downstream causal effects, and guides targeted SCM-based search.

  • Results

    76.9% reduction in active human intervention burden was achieved while maintaining global recourse cost parity and 100% relative convergence on causally feasible instances.

  • Takeaways & Limitations

    A-CBFI provides targeted causal recourse by concentrating intervention effort on diagnosed root causes while preserving structural constraints.

  • Takeaways & Limitations

    The evaluated SCMs showed limited downstream dimensional expansion, so the experiments do not support a strong burden-reduction claim from passive causal propagation alone.

Abstract

from arXiv · show

Explainable artificial intelligence (XAI) increasingly calls for actionable counterfactual recourse, yet current methodologies face challenges related to causal invalidity, excessive cognitive burden, and predictive failure. Exhaustive causal search algorithms often require modifications to multiple attributes, whereas additive attribution-guided methods, such as SHAP, ignore higher-order feature synergies, leading to suboptimal predictive momentum and diffuse intervention effort in complex nonlinear models, such as XGBoost. To bridge this gap, we introduce actionable case-based feature importance (A-CBFI), a diagnosis-prescription integrated framework for tabular machine learning. Grounded in structural causal models (SCMs), A-CBFI isolates synergistic interaction bottlenecks and releases suppressive structural locks, translating them into targeted interventions. By mathematically separating the active user intervention space (L_{\mathrm{active}}) from downstream effects and concentrating over 98.3% of the intervention effort on diagnosed root causes, A-CBFI enables highly targeted interventions. Empirical evaluations across the financial and healthcare domains demonstrate that A-CBFI reduces the active human intervention burden by 76.9% while maintaining comparable global recourse cost to exhaustive causal baselines. By prioritizing the diagnosed causal bottlenecks, A-CBFI provides targeted and actionable recourse while maintaining causal validity and achieving full relative convergence across all causally feasible instances.

1. Introduction

Existing recourse methods struggle to combine causal validity, focused intervention selection, and efficient search in nonlinear tabular models. A-CBFI addresses this gap by diagnosing synergistic bottlenecks and using them to target causally consistent interventions.

  • Motivation: Attribution methods explain predictions but do not specify feasible, causally consistent changes for improving unfavorable outcomes.This limits their direct usefulness in applications such as loan approval and clinical risk assessment.
  • Limitations of Existing Recourse: SCM-based recourse improves causal validity but exhaustive exploration can distribute modifications across peripheral attributes without prior causal prioritization.The resulting intervention space can be large and diffuse.
  • Limitations of Existing Recourse: Independent additive attributions can overlook higher-order feature combinations that dominate local decision boundaries in nonlinear models such as gradient-boosted trees.The paper calls this target-selection failure the “additive fallacy.”
  • A-CBFI Framework: A-CBFI integrates localized structural diagnosis with causal counterfactual recourse by separating diagnosis from prescription.It identifies synergistic interaction bottlenecks and uses them to restrict and prioritize intervention search within the SCM.
  • A-CBFI Framework: A-CBFI shifts recourse from exhaustive search toward targeted root-cause prescription by separating active user interventions from downstream causal effects.Interaction-Guided Beam Search prioritizes branches involving strongly interacting variables.
  • Empirical Findings: 76.9% reduction in active human intervention burden was achieved, averaging 1.72 levers while maintaining comparable global recourse cost to exhaustive baselines.The framework concentrated over 98.3% of intervention effort on diagnosed bottlenecks.

2. Methodology: A-CBFI Framework

A-CBFI converts localized CBFI decomposition into prescriptive causal guidance. It combines local main effects and higher-order synergies to diagnose structural barriers and select targeted interventions within SCMs.

  • Structural Decomposition: CBFI decomposes predictive behavior into independent, complementary, redundant, and synergistic contributions.G1 is the independent target contribution, G2 the complement contribution, G3 the shared contribution, and G4 the cooperative contribution.
  • Structural Decomposition: Localized-CBFI translates the global G1-G4 taxonomy into instance-specific diagnostics at individual decision boundaries.The localized components quantify contributions and synergies for a specific factual instance.
  • Localized Components: C_G1(x) measures the local independent contribution of the target feature set, while C_G2(x) measures the local contribution of the complement set.These components distinguish target-specific effects from effects of remaining features.
  • Localized Components: C_G3(x) captures shared or redundant effects, whereas C_G4(x) isolates high-order nonlinear synergies that can form structural barriers to recourse.C_G4 is the principal synergistic bottleneck signal used by A-CBFI.
  • From Diagnosis to Prescription: A-CBFI embeds Pearl’s intervention semantics and combines causal main effects with interaction dynamics through |C_G1|+|C_G4|.This turns localized descriptive metrics into guidance for targeted do-interventions.
  • Structural Bottlenecks: Positive C_G4 synergies can amplify adverse predictions, while negative synergistic components expose suppressive bottlenecks that targeted interventions may alleviate.The framework also retains sufficiently large adverse main effects as potential intervention targets.

2.2 Four-Step A-CBFI Pipeline

A-CBFI integrates structural causal modeling, causal diagnosis, and targeted recourse search in a four-step pipeline. It identifies actionable bottlenecks, prioritizes interaction-informed intervention paths, and preserves causal propagation while separating user actions from downstream effects.

  • Step 1: Structural Causal Model Construction: A-CBFI embeds Pearl’s do-calculus within a four-step pipeline that extends structural decomposition into actionable causal recourse.The framework replaces observational permutation attribution with causal intervention and counterfactual simulation.
  • Step 1: Structural Causal Model Construction: The SCM uses endogenous features, exogenous noise, deterministic structural equations, and an LLM-elicited DAG topology to preserve domain constraints.The topology is intended to preserve physical laws and chronological irreversibility before counterfactual simulation.
  • Step 2: Structural Causal Diagnosis via Causal CBFI: Causal diagnosis replaces observational attribution with do-operator interventions that measure main effects and nonadditive pairwise interactions toward a target outcome.C_G1 evaluates single-feature interventions, whereas C_G4 quantifies error reduction from simultaneous feature-pair interventions.
  • Step 2: Structural Causal Diagnosis via Causal CBFI: A-CBFI filters immutable attributes and retains mutable features with dominant absolute causal sensitivity as the actionable target set T.An instance-specific threshold removes negligible diffuse effects and focuses optimization on structurally impactful causal levers.
  • Step 3: Targeted Recourse Optimization via Interaction-Guided Discretized Beam Search: The targeted optimization minimizes intervention cost while balancing predictive convergence over allowable actions in T under structural causal propagation.Interaction scores guide search priorities rather than discounting the intervention cost itself.
  • Step 3: Targeted Recourse Optimization via Interaction-Guided Discretized Beam Search: Interaction-Guided Beam Search prioritizes structurally informative branches for nondifferentiable models while retaining objective intervention costs and distinguishing direct from downstream changes.The resulting search avoids exhaustive exploration of the full feature space and concentrates effort on informative intervention combinations.

2.3 Theoretical Properties of A-CBFI

A-CBFI’s theoretical account links structural and synergistic diagnosis to reduced search burden. Its interaction score serves as a signed diagnostic signal for prioritizing branches in nonlinear local decision regions.

  • Property 1: Exponential Reduction of Search Space via Structural and Synergistic Pruning: A-CBFI restricts exhaustive search from O(k^m) over all mutable features to O(k^|T|) over a diagnosed bottleneck set with |T| ≤ ⌈0.75m⌉.Interaction-guided branching further reduces unnecessary exploration within the filtered target set.
  • Property 1: Exponential Reduction of Search Space via Structural and Synergistic Pruning: The framework empirically achieves approximately 1.6–1.8 active levers on average, indicating sparse recourse interventions.This sparsity is reported alongside the claimed mitigation of the curse of dimensionality.
  • Interaction Effects and Nonlinear Decision Boundaries: Additive main-effect targeting can suffice in strictly additive models but may miss joint directions that shape nonlinear decision boundaries.The limitation is described for complex nonlinear models including XGBoost and RBF SVM.
  • Interaction Effects and Nonlinear Decision Boundaries: C_G4 preserves signed interaction magnitude to distinguish synergistic amplification from suppressive inhibition and prioritize corresponding search branches.The score is a game-theoretic diagnostic signal rather than an exact Hessian computation.

3. Experimental Setup

The evaluation uses six financial and healthcare benchmark datasets, three supervised model architectures, domain-specific SCMs, and standardized preprocessing and validation procedures. Experiments assess recourse quality, efficiency, sparsity, causal plausibility, and success across realistic structural constraints.

  • Datasets and domains: Experiments span six benchmark datasets across financial lending and healthcare diagnostics.The datasets vary in feature interactions, nonlinear boundaries, and structural constraints.
  • Datasets and domains: The financial benchmarks evaluate socioeconomic interactions, immutable demographic factors, and actionable paths under domain constraints.Financial Loan, German Credit, and Adult Income cover default risk, credit history, and exceeding the $50K income threshold.
  • Datasets and domains: The healthcare benchmarks test nonlinear physiological synergies and medically or geometrically constrained interventions.They include Medical Insurance Cost, Pima Indians Diabetes, and Breast Cancer Wisconsin.
  • Preprocessing and constraints: Continuous variables were standardized, categorical variables ordinally encoded, and selected root demographic nodes declared immutable.Immutable examples include annual_income, age, and pregnancies.
  • Evaluation protocol: The benchmark compares recourse methods using cost, intervention sparsity, computational efficiency, causal plausibility, and success rate.The evaluation samples 100 adverse test instances per dataset for counterfactual queries.
  • Predictive models: A-CBFI is evaluated with SVM, Random Forest, and XGBoost models under a standardized 80:20 train-test split.The architectures provide distinct nonlinear decision-boundary geometries and interaction structures.
  • Causal modeling: Domain-specific SCMs were constructed from domain priors, fitted from observational data, and also supplied to the untargeted causal baseline.This controls for topological advantages when assessing A-CBFI’s synergistic bottleneck diagnosis.

4. Empirical results and discussion

The empirical analysis benchmarks A-CBFI across domains, model architectures, and complementary research questions. It examines whether structural bottleneck diagnosis improves search efficiency and intervention concentration while preserving recourse quality and causal validity.

  • Evaluation scope: The benchmark spans SVM, Random Forest, and XGBoost models across six standardized datasets.This multi-architecture design tests behavior under distinct nonlinear decision boundaries.
  • Research questions: RQ1 tests whether restricting search to diagnosed causal bottlenecks reduces effort while maintaining recourse quality, causal validity, and action sparsity.The question directly evaluates the central efficiency-quality trade-off.
  • Research questions: RQ2 examines how nonlinear decision-boundary geometries influence recourse convergence and causal plausibility.The comparison is explicitly architectural rather than limited to one classifier.
  • Research questions: RQ3 asks whether interaction diagnosis identifies root-cause bottlenecks overlooked by additive attribution methods.It links diagnosis to more concentrated and structurally informed interventions.
  • Research questions: RQ4 investigates how instance-level causal recourse pathways operate dynamically within real-world SCMs.This frames actionability as a pathway-level question rather than only an outcome metric.
  • Research questions: RQ5 evaluates the contribution of synergistic diagnosis and SCM topology through ablation analysis.It asks how recourse effectiveness changes when core modules are removed.

4.1 Macroscopic Performance across Domains and Metrics

Across financial and healthcare benchmarks, A-CBFI reduces search effort and concentrates interventions while preserving comparable recourse success and SCM consistency. Its main advantage is targeted causal search rather than simply lower numerical recourse cost.

  • Overall benchmark: 26.7% fewer model evaluations were required by A-CBFI than Untargeted Causal Recourse, with success rates of 90.2% and 90.4%, respectively.Average evaluations declined from 405.0 to 296.7, while execution time fell from 1.119 s to 0.791 s.
  • Both-success comparison: 26.3% fewer evaluations were required on 1,620 jointly successful instances, comparing 320.0 for A-CBFI with 434.2 for Untargeted Causal Recourse.This controlled subset compares search effort when both causal methods successfully generate recourse.
  • Intervention concentration: A-CBFI achieved L0=1.72 versus 1.75 and RCR=98.30% versus 93.82% on the Both-Success subset.The result indicates comparable action sparsity with greater concentration of intervention effort.
  • Causal validity: SCM propagation residuals were identical at RSCM=0.0083 for both causal methods, with no statistically significant difference (p=0.554).The reduced search effort therefore did not coincide with a measured loss of SCM consistency.
  • Divergent prescriptions: 34.7% fewer evaluations and 36.6% lower runtime occurred in divergent prescriptions, while A-CBFI achieved L0=2.11 versus 2.33 and RCR=94.46% versus 63.53%.These cases involve different intervention prescriptions despite successful recourse from both causal methods.
  • Baseline comparisons: A-CBFI’s mean cost was 2.8814 and mean L0 was 1.7198, compared with 12.2365 and 7.4617 for Wachter’s CE.Relative to non-causal baselines, the framework substantially reduced cost and modified-feature count.
  • Cost and concentration: A-CBFI’s mean recourse cost was 2.8814 versus 2.7499 for Untargeted Causal Recourse, while its RCR was 98.30% versus 93.82%.Thus, A-CBFI did not minimize aggregate cost but produced more concentrated interventions.
  • Causal validity and plausibility: A-CBFI preserved causal validity comparable to Untargeted Causal Recourse and generated counterfactuals closer to the data manifold than non-causal methods.Non-causal baselines had substantially larger RSCM and DM values.

4.2 Impact of Decision Boundary Geometry on Causal Recourse Stability

Across RBF-SVM, Random Forest, and XGBoost, A-CBFI consistently reduced search effort and improved intervention precision relative to Untargeted Causal recourse. These gains persisted across smooth, axis-aligned, and adaptive piecewise decision boundaries, with modestly higher recourse cost.

  • RBF-kernel SVM: 24.8% fewer evaluations and 30.3% lower execution time were achieved by A-CBFI on the RBF-kernel SVM.Mean recourse cost was 3.094 versus 2.957, while RCR was 98.55% versus 93.02%.
  • Random Forest: 29.8% fewer evaluations and approximately 30.0% lower execution time were achieved by A-CBFI on Random Forest.A-CBFI had recourse cost 2.619 versus 2.478, RCR 97.95% versus 94.47%, and L0=1.757 versus 1.805.
  • XGBoost: 24.4% fewer objective-function evaluations and approximately 24.7% lower execution time were achieved by A-CBFI on XGBoost.A-CBFI required 330.9 versus 437.9 evaluations and 0.511 s versus 0.678 s.
  • XGBoost: On XGBoost, A-CBFI achieved RCR of 98.41% versus 93.97% and average L0 of 1.795 versus 1.841, despite recourse cost of 2.937 versus 2.818.The targeted method maintained comparable SCM consistency, with RSCM 0.005 versus 0.006.
  • Cross-architecture pattern: Across SVM, Random Forest, and XGBoost, A-CBFI consistently reduced execution time and objective-function evaluations while achieving higher RCR than Untargeted Causal recourse.Recourse cost was modestly higher across all three architectures, while causal validity remained comparable and sparsity generally improved or was preserved.

4.3 Predictive Momentum and Additive Targeting

A-CBFI was evaluated against SHAP-targeted counterfactual explanations for predictive momentum and intervention efficiency. Both methods always succeeded, but A-CBFI produced larger aggregate probability gains while SHAP-targeted CE used fewer modified features.

  • Results: Across all six domain–architecture combinations, both A-CBFI and SHAP-targeted CE achieved a 100% success rate.The comparison used absolute probability gain, intervention efficiency, and success rate.
  • Predictive momentum: A-CBFI consistently produced a larger absolute probability gain than SHAP-targeted CE across Financial and Healthcare settings.Financial ΔP values favored A-CBFI for SVM, Random Forest, and XGBoost; Healthcare showed the same pattern, including XGBoost 0.5892 versus 0.5809.
  • Intervention efficiency: SHAP-targeted CE achieved higher ΔP/L0 in every comparison by using smaller intervention sets.Thus, it generated more probability gain per modified feature despite lower absolute probability gain.
  • Interpretation: The results provide supporting rather than conclusive evidence that purely additive target selection is limited.Both methods achieved 100% success, so the evidence distinguishes predictive momentum and compactness rather than recourse success.

4.4 Qualitative Case Studies: Instance-Level Causal Pathways

The case studies show how A-CBFI converts structural diagnosis into minimal, causally consistent interventions in healthcare and finance. It separates direct user actions from downstream propagation while addressing synergistic amplification and suppressive inhibition.

  • Case-study scope: The case studies examine constrained scenarios where strong immutable constraints leave only a limited set of actionable variables.They cover synergistic amplification (C_G4 > 0) and suppressive inhibition (C_G4 < 0).
  • Healthcare case: In Diabetes Instance #152, age is immutable, leaving glucose, bmi, and pregnancies among the actionable physiological and metabolic levers.The factual adverse-risk prediction was 91.0%, with age fixed at 42 years.
  • Healthcare case: The diabetes diagnosis identified diabetespedigreefunction as interaction-dominated and exposed a pairwise bottleneck linking glucose and diabetespedigreefunction with C_G4(X,Y)=0.162.The diagnosis separated main effects from structural synergies and identified combined amplification as a major risk influence.
  • Healthcare case: A-CBFI reduced the diabetes search to diagnosed bottlenecks and used Interaction-Guided Beam Search to prioritize structurally informed intervention branches.The filtered set included glucose, pregnancies, bmi, diabetespedigreefunction, and bloodpressure.
  • Healthcare case: The diabetes prediction shifted from 91.0% adverse risk to 48.0% favorable prediction at MAD-scaled intervention cost 4.59.Only two direct levers were changed: bmi and glucose; insulin decreased passively through the SCM.
  • Financial case: In the financial case, A-CBFI prescribed one active intervention by reducing dti, while SCM propagation reduced int_rate, completing recourse in 36.84 ms.This avoided directly manipulating interest rate, which the case identifies as generally outside the applicant’s control.
  • Financial case: Wachter’s CE changed five features, while Untargeted Causal recourse directly manipulated int_rate and required 1758.06 ms.The comparisons illustrate the practical contrast between diffuse or unactionable interventions and targeted causal recourse.

4.5 Component-wise Analysis via Baseline-Induced Ablation Proxies

The baseline-induced ablation analysis maps removal of A-CBFI components to representative recourse paradigms. It attributes losses in causal plausibility, intervention burden, and recourse cost to removing SCM-based causal topology, while additive targeting weakens efficiency across nonlinear boundaries.

  • Ablation framing: The analysis is a proxy ablation rather than a strict code-level ablation, using representative baseline configurations to interpret component contributions.Table 9 maps ablated A-CBFI configurations to evaluated baseline paradigms.
  • Synergistic diagnosis: Removing synergistic diagnosis reduces A-CBFI to additive main-effect targeting analogous to SHAP-Targeted CE.This configuration maintains high success but degrades intervention efficiency across complex nonlinear decision boundaries.
  • Causal topology: Removing the SCM reduces the framework to a non-causal, distance-based paradigm represented by Wachter’s CE.This proxy ablation tests the contribution of causal topology separately from synergistic target selection.
  • Causal topology: RSCM increased from 0.0083 to 0.0737 when causal topology was removed.The same ablation increased Lactive from 1.72 to 7.46 levers and Recourse Cost from 2.8814 to 12.2365.

5. Conclusion

A-CBFI addresses the gap between theoretically valid recourse and practically manageable human actions by targeting synergistic causal bottlenecks. It reduces intervention burden while preserving causal validity and predictive effectiveness, with future work extending the framework to sequential and multimodal settings.

  • Conclusion: A-CBFI targets synergistic bottlenecks rather than distributing intervention effort across multiple mutable attributes or relying on additive SHAP explanations.The framework is designed for nonlinear models such as XGBoost, where higher-order feature synergies matter.
  • Conclusion: 98.3% of intervention momentum was concentrated on diagnosed root causes, while feasible recourse respected structural constraints and achieved 100% relative causal convergence across feasible instances.A-CBFI also maintained substantially greater probability-gain predictive momentum than SHAP.
  • Conclusion: Future studies should extend the synergistic framework to dynamic sequential recourse and multimodal data manifolds.The implementation code and experimental results are publicly available.

Appendix A: Domain-Specific SCM Priors (DOMAIN_SCM_CONFIGS)

The appendix specifies domain-specific structural causal priors, including immutable variables and directed relationships for financial, healthcare, clinical, and geometric feature systems. These priors encode mechanisms such as income, debt, age, physiology, and geometric coupling that constrain actionable recourse.

  • Financial and socio-economic domain: Financial configurations treat age, annual income, and employment length as immutable in specified datasets, while modeling dependencies among credit, borrowing, installments, interest rates, and debt burden.Examples include principal scaling monthly payment, term determining installment, and higher debt-to-income increasing interest-rate assignments.
  • Healthcare and clinical diagnostic domain: Clinical financial and diagnostic priors connect income with debt-to-income or loan amount, and employment length with interest rate, thereby representing income, borrowing, and risk relationships.These relationships supplement the healthcare and financial configurations with domain-specific lending constraints.
  • Financial and socio-economic domain: Age-based financial priors connect age with existing credits, credit amount, education, work hours, and capital gains, while duration, liabilities, and principal constrain borrowing-related features.These relationships represent life-stage, credit-history, financing-term, and labor-market mechanisms.
  • Healthcare and clinical diagnostic domain: Healthcare configurations designate age and children, or Age and Pregnancies, as immutable variables and encode physiological links involving BMI, exercise, smoking, blood pressure, glucose, insulin, and skin thickness.The listed mechanisms include aging-related stamina and metabolic changes, exercise effects on BMI, obesity-related insulin resistance, and glucose-responsive insulin secretion.
  • Clinical and geometric feature relationships: Geometric priors encode deterministic or mathematical coupling among radius, perimeter, area, compactness, texture, and smoothness features.The appendix identifies linear perimeter scaling, quadratic area scaling, and associations involving nuclear volume and surface characteristics.
Loading 2608.27821v1…