Skip to main content
Back to timeline
Communications in computer and information scienceSource publication:

Explaining a model's decision now means arguing for it: causal graphs become support-and-attack cases for why a prediction was made

Synopsis

Explainable AI methods usually name the features relevant to a prediction but leave unclear why the model reached that decision. This work joins causal discovery with argumentation: a dual-encoding FCI run estimates causal structure from observational data, positive and negative correlations become support and attack relations, and semi-stable semantics returns feature sets that coherently explain one outcome. On the Titanic and Pima Diabetes benchmarks the explanations agree with the features emphasized by SHAP and a pruned decision tree, though the authors stress the graphs are candidate causal hypotheses consistent with the data.

Source-provided article image: A Causal Argumentation Method for Explainability of Machine Learning Models
Fig. 4 · Page 11

Interpretation

Explaining a machine-learning model can go beyond listing relevant features to giving an internally consistent argument that can answer counterarguments about why this outcome rather than another. Standard post-hoc explainers rely on feature sensitivity or attribution scores, which mainly capture correlation and do not separate which features support a conclusion from which oppose it. On the two binary-classification benchmarks, Titanic and Pima Diabetes, the feature sets returned under semi-stable semantics line up with the split features of a pruned decision tree and with the top-ranked features from SHAP.

The pipeline first discretizes each feature into value ranges or categories, so that a concrete value such as age under 5, rather than the whole variable, becomes an argument. Causal structure is estimated with the FCI algorithm, which allows for unobserved confounders and outputs a partial ancestral graph, a graph where some edges are oriented and others are left undirected. Each dataset is run twice with drop-first and drop-last encodings at significance level 0.01 using Fisher's z-test, then merged by majority voting; the merged Titanic graph yields 17 arguments, 15 support relations and 17 attacks, expanding to 18 arguments and 57 attacks in the classical argumentation framework.

Perspective

The demonstration covers two binary-classification tabular datasets, Titanic (891 passengers, 7 features) and Pima Diabetes (768 female patients, 8 features), and explains discretized feature values; the authors propose future evaluation on more complex, higher-dimensional data and the introduction of quantitative metrics such as fidelity, counterfactual consistency and robustness under distribution shift.

The analysis is entirely observational, with no interventional experiments and no comparison against known ground-truth causal structures; FCI returns a Markov equivalence class, so directed edges mark orientations shared by class members and undirected edges mark unresolved direction. Both datasets are modest in size, limiting the statistical power of conditional-independence tests, and hidden confounding is plausible, such as unrecorded socioeconomic factors in Titanic and missing diet and exercise variables in Pima; faithfulness violations are also possible in finite samples.