Pith. sign in

REVIEW 4 major objections 6 minor 17 references

Generating Causally Compliant Counterfactual Explanations using ASP

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read CoGS generates counterfactual explanations as step-by-step intervention paths that respect causal dependencies among features, so a negative decision is flipped only through achievable changes.

desk verdict A useful preliminary framework that combines FOLD-SE rule learning with s(CASP) planning, but the 'realistic' claim rests on causal rules the paper never validates. read the letter →

arxiv 2502.09226 v1 pith:KZK5LLQW submitted 2025-02-13 cs.AI

Sopam Dasgupta (The University of Texas at Dallas , USA) This is my paper · ORCID
classification cs.AI
keywords counterfactualexplanationscausaldependenciesanswersetprogrammings(CASP)rule-basedmachinelearningFOLD-SEactionablerecourseplanningproblem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that counterfactual explanations—the standard 'what would need to change to flip a no into a yes'—are only trustworthy when the suggested changes respect the causal links between features. CoGS models the decision with rule-based machine learning, models feature dependencies as causal rules, and then solves a planning problem that starts at the negative outcome, ends at a positive one, and changes one feature at a time under those constraints. The payoff is an explanation that tells a user what to do first: for example, clear debt obligations so that credit score rises, rather than being told to raise the credit score directly. The paper reports preliminary results on three benchmark datasets showing that such causally ordered paths can be computed, with a dependent feature changing only after its cause has been changed.

What carries the argument

The central mechanism is a causally constrained planning problem encoded as an answer set program and executed by s(CASP), a goal-directed answer set programming system that uses program completion to turn 'if' rules into 'if and only if' rules. States are sets of feature-value pairs; each transition changes one feature value; and a causal dependency between feature values $P$ and $Q$ is encoded as both $P \Rightarrow Q$ and $\neg P \Rightarrow \neg Q$, so directly changing $Q$ is blocked unless the appropriate cause $P$ holds. FOLD-SE supplies the decision rules and the candidate causal rules, which the user approves. This machinery carries the argument because it converts 'a counterfactual the model would accept' into 'a counterfactual reachable by an ordered sequence of realistic actions'.

What would settle it

Run CoGS on a dataset whose true causal structure is known from external interventions, e.g., a loan simulator where clearing debt raises credit score but editing credit score does not change debt. If the FOLD-SE-learned rules omit the dependency or encode it backwards, CoGS's recommended interventions will fail to produce the predicted dependent-feature change or will fail to flip the decision, directly refuting the claim that the counterfactuals are realistically achievable.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that generating realistic counterfactuals for rule-based decision models can be reformulated as a planning problem over causally constrained feature states. CoGS takes decision rules from a rule-based learner, takes causal rules expressing which feature values imply (and are implied by) other feature values, and encodes both in a goal-directed answer set programming system. The system searches for an intervention path from the initial negative state to a goal state that flips the decision and is consistent with the causal rules. A dependent feature such as credit score is never directly altered; an earlier step changes its cause, such as debt obligations, and the causal rule propagates the effect. This is the paper's reason for calling the counterfactuals realistic: every step is an achievable action, and dependent changes appear as consequences of earlier steps.

Load-bearing premise

The load-bearing premise is that the rules FOLD-SE learns about feature dependencies are genuine causal laws; if they are only correlations, then the paths CoGS certifies as realistic may fail to produce the predicted changes or the promised outcome.

Editorial extensions

If this is right

  • For rule-based decision systems, CoGS returns an ordered plan of interventions, not just a target feature vector, so a user knows which change to make first.
  • Causally dependent features are changed only as effects of changing their causes; direct edits to them are excluded, and this is what makes the counterfactual achievable.
  • CoGS can enumerate all possible paths to a counterfactual and can find minimal paths by iterating over path length, so the returned explanations are both causally consistent and minimal.
  • The framework is independent of the particular rule learner: any rule-based machine learning algorithm can supply the decision and causal rules, and statistical models can be included by first approximating them with rules.
  • Preliminary experiments on German, Adult, and Car Evaluation data produced counterfactuals and paths in times of about one to three seconds, indicating that the approach is computationally feasible on small structured problems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the causal constraints come from association rules learned by FOLD-SE and verified informally, so the 'realistic' claim is only as strong as those learned dependencies; testing on externally validated causal graphs would settle how often the paths are truly achievable.
  • Beyond the paper, the planning formulation invites extensions the author does not develop here, such as attaching costs to interventions, enforcing immutability of protected features, and searching for lowest-cost rather than shortest paths.
  • Beyond the paper, one could use CoGS to generate counterfactual training examples for imbalanced datasets, as the author lists as planned work; the testable prediction is that minority-class performance improves only when the learned rules faithfully capture the decision process.
  • Beyond the paper, comparing CoGS's paths with nearest-neighbor counterfactuals on a dataset with known causal structure would quantify the realism gain and show whether the extra constrained steps are worth the added complexity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes CoGS, a framework for generating counterfactual explanations from rule-based machine learning models using Answer Set Programming (specifically s(CASP)). CoGS takes decision rules D learned by FOLD-SE, learns or accepts causal dependency rules C among features, encodes both in ASP, and solves a planning problem whose goal state is a positive outcome consistent with C and inconsistent with D. The output is a step-by-step sequence of feature-value changes from a negative outcome to a positive one. The authors argue that because these paths respect causal constraints, the resulting counterfactuals are realistic and achievable. Preliminary results are reported on three UCI datasets (German, Adult, Car Evaluation) with one example path per dataset, plus counts of counterfactual sets that CoGS can generate. The paper is presented as early-stage research with planned extensions rather than a completed evaluation.

Significance. If the central claim were established, CoGS would be a useful contribution: it combines counterfactual explanation, causal modeling, and ASP-based planning in a way that addresses a known limitation of earlier counterfactual methods that assume feature independence. The paper's strengths include a concrete, open-source implementation, a clear motivating example (John's loan application), and a sensible high-level design that separates decision rules from causal constraints and allows user approval of learned dependencies. However, the evidence offered is anecdotal: the central claim that the counterfactuals are 'realistic' is not validated against causal ground truth, interventional data, or any systematic comparison with baselines. The evaluation currently consists of three hand-selected paths, and the causal learning step is both unverified and conceptually questionable because FOLD-SE is an association-rule learner, not a causal discovery method. These issues are load-bearing for the paper's main contribution.

major comments (4)
  1. [Section 4] The paper's central claim that CoGS computes 'realistic' counterfactuals is supported only by three hand-selected example paths in Tables 1 to 3. No systematic evaluation is provided: there are no metrics for realism, actionability, cost, or validity of the generated paths, no comparison to baseline counterfactual methods such as Wachter et al., and no error analysis over the 240, 112, and 78 counterfactual sets reported in Table 4. A convincing validation would define measurable criteria (e.g., whether the suggested intermediate actions are feasible, whether the final state satisfies the decision rules, and whether the path respects known causal dependencies) and report results over all generated counterfactuals.
  2. [Section 4] The statement that 'causal rules are also learnt using FOLD-SE and verified' is not backed by any verification protocol, ground-truth causal structure, or external check. FOLD-SE is a rule-based classifier that learns association rules from observational data; such rules can hold when the causal direction is reversed, when the two features are common effects of a hidden cause, or when the dependency is definitional or selection-induced. Converting every learned association into a biconditional causal constraint, as described in Section 2, then treats spurious associations as hard causal laws. The paper must provide evidence that the learned rules correspond to genuine causal dependencies, for example by using known causal benchmarks, expert validation, or interventional data, and must report how many learned rules were accepted, rejected, or found to be spurious.
  3. [Section 2] The paper equates causality with logical equivalence: 'P is the cause of Q, if (P ⇒ Q) ∧ (¬P ⇒ ¬Q)'. This is a strong assumption that is not generally valid for causal relations, which can be asymmetric and context-dependent. Program completion in s(CASP) turns every learned 'if' rule into an 'if and only if' rule, thereby imposing bidirectionality on associations that may not be causal. This is not merely a philosophical concern: if a learned rule is spurious, CoGS may forbid an achievable intervention or force an unnecessary and unrealistic one, undermining the paper's main claim. The authors should either justify the equivalence assumption for their application domain or use a causal discovery method that produces asymmetric, directed relations.
  4. [Section 5] The paper states that 'CoGS has the ability to find minimal paths by iteratively adjusting the path length', but no formal definition of minimality is given and no proof or experiment demonstrates that the generated paths are minimal in any meaningful sense (e.g., fewest interventions, lowest cost, or most plausible). Since the paper presents minimality as a contribution, this claim needs either a precise definition with a proof or an empirical study showing that the method returns optimal or near-optimal paths according to a stated criterion.
minor comments (6)
  1. [Section 6] There is a typo in the first sentence: 'one of the limitations of of CoGS' should read 'one of the limitations of CoGS'.
  2. [Table 4 caption] The caption contains a typo: 'counterfactuals produce b by' should read 'counterfactuals produced by'. Also, the dataset names are inconsistent: 'Cars' is used in Table 4 but 'Car Evaluation' is used in the text and Table 3.
  3. [Section 2] The citation to Pearl for the definition of causality is imprecise; the formula (P ⇒ Q) ∧ (¬P ⇒ ¬Q) is more naturally a definition of logical equivalence or correlation than a Pearlian causal definition. A more careful treatment of the relationship between logical implication and causal dependence would improve the paper.
  4. [Table 2] The Adult dataset example changes 'Relationship' directly to 'husband' and then treats 'Marital Status' changing to 'married civ spouse' as a causal effect. This raises feasibility and ethical questions: a person cannot simply change their relationship status by direct action, and the path may be socially or legally constrained. The paper should discuss how such actionability issues are handled by the framework, for example through user approval of allowed interventions.
  5. [Section 4] The paper reports only successful example paths and gives no information about cases where CoGS fails to find a counterfactual, where it finds an infeasible path, or where runtime is excessive. Reporting such cases would give a more complete picture of the framework's behavior.
  6. [Section 3] The description of the goal state as 'consistent with the causal rules C and inconsistent with the decision rules D' is clear in outline, but the encoding details of the planning problem (how actions are represented, how feature domains are handled, and how numerical intervals are managed) are not given. A pseudocode listing or a small example program would help the reader understand the actual ASP implementation.

Circularity Check

1 steps flagged · score 6.0 of 10

The central 'realistic' claim reduces, by the paper's own definitions, to consistency with the FOLD-SE-learned rules; the planning computation itself is not circular.

  1. self definitional [Abstract; Section 2 (Causality Considerations); Section 3 (Research Goal)]
    "CoGS computes paths that respect the causal constraints among features. Thus, the counterfactuals computed by CoGS are realistic. ... we capture the causal dependencies C amongst the features using user-defined rules or rules learnt using RBML algorithms ... the goal state g is defined as a state that is consistent with the causal rules C and inconsistent with the decision rules D."

    The paper defines realistic counterfactuals as ones that model causal dependencies, and it operationalizes causality via s(CASP) completion that turns learned 'if' rules into 'if and only if' rules. The planning problem then defines the goal state as any state consistent with C, so every generated path is consistent with C by construction. If 'realistic' means 'respects the causal constraints,' the abstract's conclusion is a definitional tautology. If 'realistic' means 'achievable in the actual world,' the paper provides no independent evidence that the FOLD-SE-learned association rules are true causal relations—Section 4 merely says they are 'learnt using FOLD-SE and verified' without a verification protocol or causal ground truth.

full rationale

CoGS's planning computation is not circular: given a set of causal constraints, finding a path to a goal that respects those constraints is a genuine ASP planning problem, and the runtime results in Tables 1-4 are real computational outputs. The circularity is confined to the paper's central 'realistic' claim. The abstract infers 'realistic' directly from 'respect the causal constraints among features.' In Section 2, causality is operationalized as completion of learned if-rules into iff-rules, and in Section 3 the goal state is defined as a state consistent with those causal rules C. Hence every generated counterfactual satisfies C by construction. If 'realistic' means 'respects C,' the claim is a tautology; if it means 'achievable in the actual world,' the paper supplies no evidence that the FOLD-SE-learned association rules are true causal relations—Section 4 only says they are 'learnt using FOLD-SE and verified' with no protocol or ground truth. Thus the load-bearing prediction reduces to the fitted/completed input C. There is no significant load-bearing self-citation (the GitHub supplement is auxiliary), so the score is moderate rather than maximal.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The framework rests on assumptions about how causality can be represented in ASP and learned from data. No explicit numeric free parameters are reported, but the learned rule sets themselves are fitted models and carry the main uncertainty.

assumptions (4)
  • domain assumption s(CASP) program completion turns 'if' rules into 'iff' rules, and this models causality.
    Section 2 states that completion models causality, which is a strong modeling assumption that equates logical biconditionals with causal dependencies.
  • ad hoc to paper FOLD-SE can discover causal dependencies between features from observational data.
    Section 4 says causal rules are learned by FOLD-SE and verified, but FOLD-SE is a rule-based classifier, not a causal discovery method; this assumption is not independently justified.
  • domain assumption The decision rules D and causal rules C correctly represent the domain, possibly with user approval.
    Section 3 defines the goal state as consistent with C and inconsistent with D; the framework inherits any errors in these rule sets.
  • domain assumption Each state transition changes exactly one feature value, and causal constraints permit only certain transitions.
    Section 1 describes the path as a sequence of nodes where each node changes an attribute value; this atomic-intervention model is assumed throughout.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generating Causally Compliant Counterfactual Explanations using ASP." pith.science (2026). https://pith.science/paper/KZK5LLQW

@misc{pith2026250209226,
  author       = {Pith},
  title        = {Pith review of: Generating Causally Compliant Counterfactual Explanations using ASP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KZK5LLQW}},
  note         = {Machine review of arXiv:2502.09226}
}
read the original abstract

This research is focused on generating achievable counterfactual explanations. Given a negative outcome computed by a machine learning model or a decision system, the novel CoGS approach generates (i) a counterfactual solution that represents a positive outcome and (ii) a path that will take us from the negative outcome to the positive one, where each node in the path represents a change in an attribute (feature) value. CoGS computes paths that respect the causal constraints among features. Thus, the counterfactuals computed by CoGS are realistic. CoGS utilizes rule-based machine learning algorithms to model causal dependencies between features. The paper discusses the current status of the research and the preliminary results obtained.

Figures

Figures reproduced from arXiv: 2502.09226 by the authors.

Figure 1
Figure 1. Top: Example 1 shows how John goes from being rejected for a loan to having his loan approved. Here the bank only considers the bank balance for loan approval. John does a direct action to increase his bank balance to $60000. Bottom: Example 2 shows how John goes from being rejected for a loan to having his loan approved. Here the bank considers both bank balance as well as credit score for loan approval. While the … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages

  1. [1]

    Theory Pract

    Joaqu´ ın Arias, Manuel Carro, Elmer Salazar, Kyle Marpl e & Gopal Gupta (2018): Constraint Answer Set Programming without Grounding . Theory Pract. Log. Program. 18(3-4), pp. 337–354, doi: 10.1017/ S1471068418000285

    work page 2018

  2. [2]

    Cambridge University Press, doi:10.1017/CBO9780511543357

    Chitta Baral (2003): Knowledge Representation, Reasoning and Declarative Prob lem Solving . Cambridge University Press, doi:10.1017/CBO9780511543357

    work page doi:10.1017/cbo9780511543357 2003

  3. [3]

    UCI Machine Learning Repository

    Barry Becker & Ronny Kohavi (1996): Adult. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5XW20

    doi:10.24432/c5xw20 1996

  4. [4]

    Bertossi (2020): An ASP-Based Approach to Counterfactual Explanations for C lassifica- tion

    Leopoldo E. Bertossi (2020): An ASP-Based Approach to Counterfactual Explanations for C lassifica- tion. In V´ ıctor Guti´ errez-Basulto, Tom´ as Kliegr, Ahmet Soylu, Martin Giese & Dumitru Roman, editors: Rules and Reasoning - 4th International Joint Conference, R uleML+RR 2020, Oslo, Norway, June 29 - July 1, 2020, Proceedings , Lecture Notes in Computer...

    work page 2020

  5. [5]

    UCI Machine Learning Repository

    Marko Bohanec (1997): Car Evaluation . UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5JP48

    doi:10.24432/c5jp48 1997

  6. [6]

    Gerhard Brewka, Thomas Eiter & Miroslaw Truszczynski (2 011): Answer set programming at a glance . Commun. ACM 54(12), pp. 92–103, doi: 10.1145/2043174.2043195

    work page arXiv

  7. [7]

    Available at https://github.com/sopam/Supplementary

    Sopam Dasgupta (2024): Supplement: CoGS: Causality Constrained Counterfactual E xplanations using Goal-directed ASP. Available at https://github.com/sopam/Supplementary

    work page 2024

  8. [8]

    Gelfond & Y

    M. Gelfond & Y . Kahl (2014): Knowledge representation, reasoning, and the design of int elligent agents: Answer Set Programming approach. Cambridge Univ. Press, doi: 10.1017/CBO9781139342124

    doi:10.1017/cbo9781139342124 2014

Show all 17 references
  1. [9]

    UCI Machine Learning Repository

    Hans Hofmann (1994): Statlog (German Credit Data) . UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5NC77

    1994 doi

  2. [10]

    Amir-Hossein Karimi, Gilles Barthe, Borja Balle & Isab el V alera (2020): Model-Agnostic Counterfactual Explanations for Consequential Decisions . In Silvia Chiappa & Roberto Calandra, editors: The 23rd In- ternational Conference on Artificial Intelligence and Stat istics, AIST...

    2020

  3. [11]

    In: Proc

    Amir-Hossein Karimi, Bernhard Sch¨ olkopf & Isabel V alera (2021): Algorithmic Recourse: from Counterfac- tual Explanations to Interventions . In: Proc. ACM FAccT, pp. 353–362, doi: 10.1145/3442188.3445899

    2021

  4. [12]

    Available at https://archive.ics.uci.edu/

    Kolby Nottingham Markelle Kelly, Rachel Longjohn: UCI Machine Learning repository . Available at https://archive.ics.uci.edu/

  5. [13]

    Statistics Surveys 3(none), pp

    Judea Pearl (2009): Causal inference in statistics: An overview . Statistics Surveys 3(none), pp. 96 – 146, doi:10.1214/09-SS057

    2009 doi

  6. [14]

    In danah boyd & Jamie H

    Berk Ustun, Alexander Spangher & Y ang Liu (2019): Actionable Recourse in Linear Classification . In danah boyd & Jamie H. Morgenstern, editors: Proceedings of the Conference on Fairness, Accountability , and Transparency, FA T* 2019, Atlanta, GA, USA, January 29-3 1, 2019 , AC...

    2019

  7. [15]

    Mittelstadt & Chris Russell (2 017): Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR

    Sandra Wachter, Brent D. Mittelstadt & Chris Russell (2 017): Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR . CoRR abs/1711.00399. arXiv: 1711.00399

    arXiv

  8. [16]

    Huaduo Wang & Gopal Gupta (2024): FOLD-SE: An Efficient Rule-Based Machine Learning Algorith m with Scalable Explainability. In Martin Gebser & Ilya Sergey, editors: Practical Aspects of Declarative Languages - 26th International Symposium, P ADL 2024, London, UK, January 15-16...

    2024 doi

  9. [17]

    d’Avila Garcez (2020): Measurable Counterfactual Local Explanations for Any Classifier

    Adam White & Artur S. d’Avila Garcez (2020): Measurable Counterfactual Local Explanations for Any Classifier . In Giuseppe De Giacomo, Alejandro Catal´ a, Bistra Dilkina, Michela Milano, Sen´ en Barro, Alberto Bugar´ ın & J´ erˆ ome Lang, editors:ECAI 2020 - 24th European Confe...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.