A Causal Argumentation Method for Explainability of Machine Learning Models

Scientific Impact Assessment

Core Contribution

This paper proposes a method that combines causal discovery (using the FCI algorithm with a dual-encoding strategy) with Bipolar Argumentation Frameworks (BAFs) to generate explanations for machine learning model predictions. The key idea is to: (1) discover causal relationships among features via constraint-based methods, (2) translate the causal graph into a BAF where positive/negative correlations define support/attack relations, (3) convert the BAF to a classical argumentation framework, and (4) compute semi-stable extensions that represent coherent sets of features supporting or opposing a given outcome.

The contribution sits at the intersection of causal discovery, formal argumentation, and explainable AI. The authors differentiate their work from Bistarelli et al. (2022) by integrating causal discovery algorithms rather than relying solely on correlation-based argumentation structures.

Methodological Rigor

The methodology has several concerning gaps:

Evaluation weakness. The most significant limitation is the absence of quantitative evaluation metrics. The paper validates its approach by qualitatively comparing semi-stable extensions against SHAP values and pruned decision trees on two small benchmark datasets (Titanic: 891 samples, Pima Diabetes: 768 samples). The comparison is purely visual and narrative — the authors observe that similar features appear as important across methods and declare "convergent validity." This falls well short of rigorous evaluation. There are no fidelity metrics, no formal agreement scores, no statistical tests, and no user studies to assess whether the explanations are actually more interpretable or useful than existing methods.

Disconnect between model and explanations. A fundamental concern is that the method does not appear to explain any specific trained ML model's behavior. The causal discovery is performed on the data itself, and the argumentation framework is built from causal/correlational relationships in the data. The resulting extensions describe coherent feature profiles associated with outcomes in the data, not necessarily what a particular model has learned. This is a significant conceptual gap — the method explains data patterns rather than model decisions, which undermines its positioning as an XAI technique.

Probabilistic constellation approach. The paper mentions generating the "top k=50 possible worlds" for probabilistic reasoning over attack relations, but provides almost no detail on how these worlds are constructed, sampled, or aggregated. This is a critical methodological component left underspecified.

Causal discovery limitations acknowledged but not addressed. The authors commendably discuss limitations of observational causal discovery (Section 5), including hidden confounding, faithfulness violations, and equivalence class ambiguity. However, these limitations fundamentally affect the reliability of the entire pipeline, and no sensitivity analysis or robustness checks are provided.

Potential Impact

The idea of combining causal reasoning with argumentation for explainability is intellectually appealing and addresses a genuine gap — most XAI methods provide feature importance scores without explaining the reasoning structure behind predictions. The argumentation framework naturally captures the dialectical nature of competing explanations, which aligns with how humans reason.

However, the practical impact is limited by several factors:

Timeliness & Relevance

The paper addresses a timely topic — the need for deeper, more structured explanations beyond feature attribution scores. The integration of causality into XAI is an active research direction, and formal argumentation provides a principled reasoning framework. However, the execution does not yet meet the bar needed to advance the state of the art convincingly.

Strengths

1. Conceptual appeal: The marriage of causal discovery and argumentation theory for explainability is a creative and well-motivated idea.

2. Theoretical grounding: The choice of semi-stable semantics is well-justified with clear reasoning about why it suits the explainability context.

3. Dual-encoding strategy: The approach to handling categorical variables in causal discovery (addressing covariance matrix singularity) is a practical contribution.

4. Honest limitations section: The authors provide a thorough and transparent discussion of the limitations of their causal claims.

5. Code availability: The implementation is publicly available.

Limitations

1. No quantitative evaluation: The comparison with SHAP and decision trees is entirely qualitative and subjective.

2. Model-agnosticism is illusory: The method doesn't actually explain a model — it explains data patterns, which is fundamentally different from what SHAP and LIME do.

3. Only two small, well-known datasets: These are insufficient to demonstrate generalizability or scalability.

4. Underspecified components: The probabilistic constellation approach and threshold selection (τ) are not adequately described.

5. No comparison with related causal XAI methods: The paper doesn't compare against other causal explanation methods (e.g., causal SHAP, CXPlain, or other causal XAI approaches).

6. The extensions produced are quite generic: The Titanic extensions essentially say "males in third class died; females in first class survived" — information trivially obtainable from basic data exploration.

Overall Assessment

This paper presents an interesting conceptual framework but falls short in execution. The lack of quantitative evaluation, the disconnect between explaining data versus explaining models, the restriction to small benchmark datasets, and underspecified methodological components significantly limit its impact. The paper reads more as a preliminary proof-of-concept than a mature contribution ready to influence the field. The honest limitations discussion is appreciated but highlights rather than resolves the fundamental issues.