Automated Explanation Selection for Scientific Discovery

Automated reasoning is a key technology in the young but rapidly growing field of Explainable Artificial Intelligence (XAI). Explanability helps build trust in artificial intelligence systems beyond their mere predictive accuracy and robustness. In this paper, we propose a cycle of scientific discovery that combines machine learning with automated reasoning for the generation and the selection of explanations. We present a taxonomy of explanation selection problems that draws on insights from sociology and cognitive science. These selection criteria subsume existing notions and extend them with new properties.

💡 Research Summary

**
The paper “Automated Explanation Selection for Scientific Discovery” proposes a novel scientific discovery cycle that tightly integrates machine learning (ML) with automated reasoning (AR) to generate, evaluate, and select explanations for AI models. The authors argue that trustworthiness in high‑risk AI applications (e.g., healthcare, autonomous vehicles, law enforcement) requires more than predictive accuracy; it demands intelligible, verifiable rationales that can be inspected, audited, and used to formulate testable scientific hypotheses.

The proposed cycle consists of three phases. In the inductive phase, a conventional ML model is trained on existing data. In the deductive phase, the trained model is encoded into a formal language (typically propositional SAT, SMT, or related extensions) and an AR engine is used to derive logical explanations. Because modern SAT/SMT solvers produce proof objects (certificates of satisfiability or unsatisfiability), the derived explanations come with mathematically rigorous guarantees that can be checked by formally verified software. The third phase, explanation selection, is where human researchers compare the formally derived explanations with prior domain knowledge, select those that are most useful, and turn them into hypotheses that can be experimentally tested. The results of new experiments feed back into the inductive phase, closing the loop.

A substantial portion of the paper is devoted to justifying why SAT‑based reasoning is especially suitable for this task. The authors highlight three properties: (1) efficiency – recent advances in CDCL, portfolio, and parallel SAT solving (e.g., Mallob) make it feasible to solve NP‑hard explanation‑selection instances; (2) trustworthiness – solvers emit checkable proof traces, allowing independent verification of every deduction; (3) genericity – by altering the encoding of the problem rather than the algorithm, the same infrastructure can address a wide variety of explanation‑selection variants (e.g., AllSAT enumeration, MaxSAT optimization, weighted variants).

The core contribution lies in a taxonomy of “explanation‑selection problems” that draws on decades of research from philosophy, psychology, sociology, and cognitive science. The authors identify eight desiderata that human users typically value in explanations:

1. Contrastiveness – explanations answer “Why P instead of Q?” (the foil or contrast case).
2. Necessity – the selected features must be required for the outcome.
3. Sufficiency – the selected features must be enough to guarantee the outcome.
4. Minimality – explanations should be as short as possible (subset‑minimal or cardinality‑minimal).
5. Generality – explanations that apply across many situations are preferred.
6. Anomaly – unusual events are often chosen as explanatory factors.
7. Probability – the most likely explanations are favored when multiple candidates exist.
8. Statistical Generalization – explanations that can be linked to broader statistical patterns need an additional causal narrative.

These criteria are not independent; they often conflict (e.g., a highly general explanation may be less minimal). The authors therefore formulate explanation selection as a multi‑objective optimization problem, most naturally expressed as a MaxSAT or weighted MaxSAT instance. By assigning weights to each desideratum, the solver can enumerate explanations that best trade‑off the competing goals, or enforce hard thresholds to prune the search space.

The paper distinguishes two formal explanation families that have recently become central in the XAI community: Abductive explanations (AXp), also known as prime implicants or minimally sufficient reasons, and Contrastive explanations (CXp), also called minimal required changes. AXp captures a minimal set of features that guarantee a prediction, while CXp captures a minimal set of feature changes that would flip the prediction. The authors emphasize the duality between AXp and CXp: CXp corresponds to a minimal hitting set of AXp explanations and vice‑versa. This duality links explanation enumeration to classic logic problems such as Minimum Unsatisfiable Subsets (MUS) and Minimum Correction Sets (MCS), which are Σ₂^P‑hard. Nevertheless, recent SAT‑based algorithms (e.g., MARCO) can enumerate these sets efficiently, especially when combined with modern clause‑learning heuristics and symmetry‑breaking techniques.

A concrete application domain discussed is tree ensembles (random forests, gradient‑boosted trees). Although each individual tree is interpretable, the ensemble as a whole is not. The authors show that by encoding the ensemble’s decision logic into SAT/SMT, exact AXp and CXp can be computed in practice, outperforming heuristic methods such as LIME or Anchors both in runtime and in guaranteeing minimality. They also mention a hybrid system, Asteryx, which first builds a surrogate random forest via sampling and then applies formal explanation techniques to the surrogate, thereby extending the approach to black‑box neural networks.

In the final sections, the authors argue that the provable guarantees of formal explanations simplify evaluation. Because each explanation is accompanied by a proof, the usual subjectivity of human‑centered XAI evaluation (e.g., user studies, surveys) can be reduced. Moreover, regulatory frameworks such as the European AI Act demand demonstrable transparency; formally verified explanations can serve as legally admissible evidence of compliance.

Overall, the paper presents a compelling vision: a fully automated, mathematically rigorous pipeline that turns opaque ML models into a curated set of human‑friendly explanations, which can then be used as scientific hypotheses, compliance artifacts, or decision‑support tools. By bridging automated reasoning, formal logic, and social‑science insights on explanation quality, the work lays a solid foundation for future research on trustworthy, explainable AI in high‑stakes scientific and societal contexts.