pith. sign in

arxiv: 2606.24348 · v1 · pith:LJS3DOYNnew · submitted 2026-06-23 · 💻 cs.CE · stat.ML

A Step Towards Inherently Interpretable Causal Machine Learning Models For Decision Support

David Zapata Gonzalez This is my paper

Pith reviewed 2026-06-25 21:59 UTC · model grok-4.3

classification 💻 cs.CE stat.ML
keywords causal machine learninginherently interpretable modelscross-sectional datawhat-if analysisdecision supportcausal relationshipsmodel transparencypredictive accuracy
0
0 comments X

The pith

Integrating causal machine learning with inherently interpretable models yields competitive prediction and what-if performance plus transparency on structure, causal links, and functional forms for cross-sectional data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes combining causal machine learning methods with inherently interpretable models to handle cross-sectional data in decision support settings. This combination is evaluated for its ability to deliver accurate predictions and support what-if scenario analysis while exposing the underlying system structure. A sympathetic reader would care because many practical decisions depend on causal understanding rather than correlations alone, allowing users to assess interventions transparently. The work positions this integration as a way to produce models that are both useful for counterfactual reasoning and open about their internal causal relationships and functional connections.

Core claim

The paper claims that the proposed integration of causal machine learning with inherently interpretable models for cross-sectional data achieves competitive performance in prediction and what-if analysis while offering transparency on the system structure, causal relationships among variables, and the functional forms that connect them. This enables informed, transparent, and causally grounded decisions and contributes to research on causality, machine learning interpretability, and data-driven decision support.

What carries the argument

The integration of causal machine learning with inherently interpretable models for cross-sectional data, which simultaneously supports predictive accuracy, interventional what-if queries, and explicit transparency on causal structure and functional forms.

If this is right

  • The models support what-if scenario evaluation grounded in causal relationships.
  • Users gain direct visibility into system structure, variable causal links, and connecting functional forms.
  • Predictive performance remains competitive with non-causal approaches.
  • The method advances data-driven decision support that is both accurate and structurally transparent.
  • It bridges causality research with interpretability needs in applied sectors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same integration strategy could be tested on longitudinal or time-series data to check whether the transparency benefits persist.
  • Widespread adoption might reduce dependence on separate post-hoc explanation tools when causal reasoning is required.
  • Domain-specific trials in areas such as medical or policy decisions could reveal whether the explicit causal forms improve trust and outcome quality.
  • Scaling the approach to high-dimensional inputs would test whether the claimed lack of trade-offs holds beyond the evaluated settings.

Load-bearing premise

Causal machine learning methods can be combined with inherently interpretable models for cross-sectional data without trade-offs that undermine predictive accuracy, what-if capability, or structural transparency.

What would settle it

A controlled test on data with known causal structure where the integrated model either loses significant predictive accuracy relative to standard machine learning baselines or produces counterfactual predictions that diverge from the true causal effects.

read the original abstract

The growing reliance on machine learning for decisions across sectors underscores the importance of model transparency and interpretability. Existing post hoc explainability methods and inherently interpretable approaches shed light on model behavior, yet they primarily reveal how models exploit correlations to maximize performance in prediction tasks. However, many decisions require causal insights and the possibility of using models for what-if scenario evaluation. To address this, we propose the integration of causal machine learning with inherently interpretable models for cross-sectional data. We evaluate these methods in terms of predictive accuracy and interpretability. Our findings show that the proposed approach achieves competitive performance in prediction and what-if analysis while offering transparency on the system structure, causal relationships among variables, and the functional forms that connect them. This work contributes to research on causality, machine learning interpretability, and data-driven decision support by offering informed, transparent, and causally grounded decisions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes integrating causal machine learning methods with inherently interpretable models for cross-sectional data. It evaluates the integration in terms of predictive accuracy and interpretability, claiming that the approach achieves competitive performance in both prediction and what-if analysis while providing transparency on system structure, causal relationships among variables, and the functional forms connecting them.

Significance. If the integration can be shown to simultaneously deliver competitive predictive performance, reliable counterfactuals, and structural transparency without trade-offs, the work would contribute to the development of decision-support models that combine causal grounding with interpretability. This addresses a recognized gap between correlation-driven ML and the needs of what-if decision tasks.

major comments (2)
  1. [Abstract] Abstract: The central claim that the proposed approach 'achieves competitive performance in prediction and what-if analysis' while offering transparency without trade-offs is unsupported by any quantitative results, baseline comparisons, held-out intervention metrics, or experimental protocol. This absence is load-bearing because the manuscript's contribution rests entirely on the asserted outcome of the integration.
  2. [Abstract] Abstract: No description is given of the specific causal ML techniques, the choice of inherently interpretable model class, the cross-sectional datasets, or the procedure used to extract and validate causal structure and functional forms. Without these elements the feasibility of the claimed combination cannot be assessed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their review and for highlighting issues with the abstract. We agree that the abstract as written does not sufficiently detail the methods, data, or supporting results, and we will revise it to address these points directly while preserving the manuscript's core contribution.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the proposed approach 'achieves competitive performance in prediction and what-if analysis' while offering transparency without trade-offs is unsupported by any quantitative results, baseline comparisons, held-out intervention metrics, or experimental protocol. This absence is load-bearing because the manuscript's contribution rests entirely on the asserted outcome of the integration.

    Authors: We accept that the abstract states the performance claim without referencing specific quantitative evidence or protocols. The body of the manuscript reports experimental results on predictive accuracy, what-if analysis, and interpretability metrics across cross-sectional datasets with baseline comparisons. To resolve the concern, we will revise the abstract to include a concise summary of the experimental protocol, key metrics, and main findings so that the claim is grounded in the abstract itself. revision: yes

  2. Referee: [Abstract] Abstract: No description is given of the specific causal ML techniques, the choice of inherently interpretable model class, the cross-sectional datasets, or the procedure used to extract and validate causal structure and functional forms. Without these elements the feasibility of the claimed combination cannot be assessed.

    Authors: The abstract is written at a high level. Specific causal ML techniques, the interpretable model class, the datasets employed, and the procedures for extracting and validating causal structures and functional forms are described in the Methods and Experiments sections. We will revise the abstract to incorporate brief, concrete references to these elements (e.g., naming the causal method family, model class, and dataset characteristics) so that feasibility can be assessed from the abstract. revision: yes

Circularity Check

0 steps flagged

No derivation chain or equations present; empirical claims not reducible by construction

full rationale

The provided abstract and description contain no equations, derivations, fitted parameters presented as predictions, or self-citations used as load-bearing premises. The central claims concern empirical evaluation of an integration approach for predictive accuracy and interpretability, which are presented as outcomes of testing rather than mathematical identities or self-referential definitions. No steps match the enumerated circularity patterns, as there is no visible reduction of results to inputs by construction. This is the expected outcome for a methods proposal paper lacking explicit formal derivations.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no concrete free parameters, axioms, or invented entities; ledger left empty.

pith-pipeline@v0.9.1-grok · 5673 in / 1022 out tokens · 23989 ms · 2026-06-25T21:59:27.022376+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    What is the projected sales volume next quarter given current market trends?

    Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 1 A STEP TOWARDS INHERENTLY INTERPRETABLE CAUSAL MACHINE LEARNING MODELS FOR DECISION SUPPORT Completed Research Paper David Zapata Gonzalez, Paderborn University, Paderborn, Germany, david.zapata@upb.de Abstract “The growing reliance on machine learning for decisions acros...

    2026

  2. [2]

    In the example of a chain, conditioning on Y blocks the relationships between X and Z; in a fork, adjusting for the common cause X removes spurious association between Y and Z; and in a collider, conditioning on the shared effect Z introduces bias by creating a spurious correlation between otherwise independent variables X and Y (Molak, 2023). Figure

    2023

  3. [3]

    Variables can interact in complex networks that involve many interconnected nodes, but the fundamental principles underlying chains, forks, and colliders remain consistent

    Basic structures in causal graphs. Variables can interact in complex networks that involve many interconnected nodes, but the fundamental principles underlying chains, forks, and colliders remain consistent. Conditional independence, through the idea of blocking paths, serves as a key criterion for determining when it is appropriate to control for certain...

    2017

  4. [4]

    Y” with two predictors “X1

    and Local Interpretable Model-agnostic Explanations (LIME) (Ribeiro et al., 2016). In contrast, inherently interpretable models are designed so that their internal logic is transparent and directly accessible to humans (Rudin, 2019). In this regard, IIML models provide transparency by design, ensuring that their recommendations are directly understandable...

    2016

  5. [5]

    Partial dependence plots and equations for inherently interpretable models. Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 6 The following four subsections provide a more detailed explanation of each part of the framework, and the fifth describes the evaluation procedure....

    2026

  6. [6]

    A combination of causal graphs with IIML models

    Proposed framework. A combination of causal graphs with IIML models. Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 7 In summary, there are tools available to support the extraction and verification of causal graphs, but ultimately, it is not possible to confirm the exact...

    2026

  7. [7]

    sin", "log

    Two modelling approaches for cross-sectional data. Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 8 3.3 Possible Task In purely predictive tasks, incorporating all available features into a machine learning model, even those that are only spuriously correlated with the ta...

    2026

  8. [8]

    Figure 6 (a) depicts the first-level, association post-hoc interpretation, by using summary SHAP values of the XGBoost

    Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 10 Type Model MAE RMSE WAPE Type Model MAE RMSE WAPE CML SR 161.7618 201.7267 0.1381 CML SR 546.3665 685.4107 0.3378 CML GAM 172.7371 221.7026 0.1474 CML GAM 655.6439 818.8543 0.4054 ML GAM 258.9799 328.4655 0.2210 CML MLP 96...

    2026

  9. [9]

    (a) SHAP summary for the xXGBoost; (b) equation found by the SR; (c) GAM´s PDPs

    Post-hoc and inherent interpretation associations. (a) SHAP summary for the xXGBoost; (b) equation found by the SR; (c) GAM´s PDPs. (b) 𝑒𝑛𝑒𝑟𝑔𝑦=−5526.7470+(𝑝𝑒𝑟𝑠𝑜𝑛𝑛𝑒𝑙·993.9957) + (𝑝𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦 · 132.6140) (c) (a) Table

    arXiv

  10. [10]

    Results for the test set. Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 11 assign substantial importance to the personnel variable, even though it has no causal link to the target outcome. Due to confounding with size, the models automatically selected personnel based on...

    2026

  11. [11]

    (a) Causal graph with SR equations; (b) GAM PDPs

    Causal inherent interpretation. (a) Causal graph with SR equations; (b) GAM PDPs. = 63,5110 + (ln (1+material) + 1,7779/3,1755) · size =9998,7280 + (productivity · 130,0056) + size2 =14.9813 0.001 · size2 (b) (a) Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 12 4.2 Evalu...

    2026

  12. [12]

    Results for the interventions. Inherently Interpretable Causal Machine Learning Thirty-Fourth European Conference on Information Systems (ECIS 2026), Milan, Italy 13 In the proposed framework and demonstration, we distinguished between several levels of interpretability: association post-hoc interpretation, association inherent interpretation, causal post...

    2026

  13. [13]

    Future research should examine how such approaches can be integrated with IIML models

    and meta-learners (Künzel et al., 2019). Future research should examine how such approaches can be integrated with IIML models. Also, in this study, we did not conduct a detailed analysis of why the different models performed as they did, for example, why the IIML models achieved such strong results, while traditional black-box models for tabular data, su...

    work page doi:10.1007/s12525-021-00459-2 2019

  14. [14]

    https://doi.org/10.3389/fgene.2019.00524 Hasan, U., Hossain, E., & Gani, M. O. (2024). A Survey on Causal Discovery Methods for I.I.D. and Time Series Data (arXiv:2303.15027). arXiv. http://arxiv.org/abs/2303.15027 Hastie, T., & Tibshirani, R. (1986). Generalized additive models. Statistical Science, 1(3), 297–310. Hastie, T., Tibshirani, R., & Friedman, ...

    work page doi:10.3389/fgene.2019.00524 2019

  15. [15]

    Machlev, L

    Machlev, R., Heistrene, L., Perl, M., Levy, K. Y., Belikov, J., Mannor, S., & Levron, Y. (2022). Explainable Artificial Intelligence (XAI) techniques for energy and power systems: Review, challenges and opportunities. Energy and AI, 9, 100169. https://doi.org/10.1016/j.egyai.2022.100169 Makke, N., & Chawla, S. (2024). Interpretable scientific discovery wi...

    work page doi:10.1016/j.egyai.2022.100169 2022

  16. [16]

    https://doi.org/10.1007/s10462-023-10622-0 Meske, C., Bunde, E., Schneider, J., & Gersch, M. (2022). Explainable Artificial Intelligence: Objectives, Stakeholders, and Future Research Opportunities. Information Systems Management, 39(1), 53–63. https://doi.org/10.1080/10580530.2020.1849465 Molak, A. (2023). Causal Inference and Discovery in Python: Unlock...

    work page doi:10.1007/s10462-023-10622-0 2022

  17. [17]

    "Why Should I Trust You?": Explaining the Predictions of Any Classifier

    https://doi.org/10.1038/s41467-021-23479-0 Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?”: Explaining the Predictions of Any Classifier (arXiv:1602.04938). arXiv. https://doi.org/10.48550/arXiv.1602.04938 Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41467-021-23479-0 2016

  18. [18]

    Zschech, P., Weinzierl, S., & Kraus, M. (2025). Inherently Interpretable Machine Learning: A Contrasting Paradigm to Post-hoc Explainable AI. Business & Information Systems Engineering. https://doi.org/10.1007/s12599-025-00964-0

    work page doi:10.1007/s12599-025-00964-0 2025