BoolXLLM: LLM-Assisted Explainability for Boolean Models

Technical Summary: BOOLXLLM

Problem Statement
While interpretable machine learning aims to provide transparent models, practical deployment in high-stakes domains (finance, healthcare) faces two persistent challenges when using expressive Boolean rule-based models like BOOLXAI. First, in high-dimensional datasets, practitioners must manually select semantically meaningful features and determine how to discretize numerical variables into meaningful categories; these decisions rely heavily on domain expertise and significantly impact model interpretability. Second, even when Boolean rules are structurally compact, they often appear abstract to non-technical stakeholders. A rule such as Or(duration > 393, nr.employed <= 5076, month = mar) lacks semantic grounding without additional context, hindering its accessibility to human decision-makers.

Methodology: The BOOLXLLM Framework
The paper proposes BOOLXLLM, a hybrid framework that integrates Large Language Models (LLMs) into the end-to-end pipeline of Boolean rule learning. This approach augments the existing BOOLXAI classifier—which learns compact logical rules using operators like AtLeast, AtMost, and Choose—with LLMs at three critical stages to bridge the gap between formal logic and human narratives:

1. LLM-Guided Feature Selection:
   Instead of relying solely on statistical correlation, the framework uses LLMs as a semantic interface between dataset metadata and feature selection. By providing the LLM with the prediction objective, feature names, descriptions, and optional domain context, the model identifies and ranks features based on semantic relevance. The output is constrained to a structured JSON format to ensure deterministic parsing, effectively reducing the search space while preserving domain-relevant variables.

2. LLM-Based Threshold Recommendation:
   Numerical features in Boolean rule learning require discretization. Standard binning often yields arbitrary splits (e.g., age > 36.33) that lack interpretability. BOOLXLLM formulates threshold selection as a structured prompting task where the LLM proposes context-aware discretization thresholds aligned with domain semantics (e.g., age > 35). These thresholds are generated prior to the rule optimization process, ensuring the resulting conditions align with human reasoning.

3. Rule Compression and Interpretation:
   To address the abstraction of Boolean rules, the framework employs LLMs for both global and local interpretation:

   • Global Level: Boolean expressions are first translated into natural language descriptions. These texts are embedded and clustered to identify "rule families"—structurally different but semantically similar rules. The LLM then summarizes these clusters to reveal population-level patterns and behavioral segments.
   • Local Level: The LLM generates instance-specific explanations by conditioning on both the learned Boolean rule and the specific input sample, translating the formal logic into natural language narratives.

Crucially, the framework grounds all LLM-generated explanations in the explicit Boolean rules learned by BOOLXAI. This design mitigates the risk of hallucination and ensures that the linguistic output remains faithful to the underlying model's decision process.

Key Contributions

• Novel Framework: Introduction of BOOLXLLM, which integrates LLMs into interpretable rule-based learning for feature selection, thresholding, and rule interpretation.
• Semantic Grounding: Methods for LLM-guided feature selection and threshold recommendation that improve the semantic relevance of Boolean rules.
• Human-Centric Translation: Development of global and local interpretation techniques that translate formal logic into accessible natural language.
• Empirical Validation: Demonstration that LLM-assisted pipelines can enhance interpretability without compromising predictive performance.

Experimental Results
The authors evaluated BOOLXLLM on the UCI Bank Marketing dataset, comparing it against a baseline BOOLXAI model using all features and standard discretization.

• Feature Selection: Using LLM guidance, the system successfully filtered 4 out of 20 features (a 20% reduction) based on semantic relevance, achieving a balanced accuracy of 86%, comparable to the baseline using all features.
• Thresholding: LLM-generated thresholds (e.g., duration > 400) aligned better with domain semantics than standard binning (e.g., duration > 393) without any loss in predictive power.
• Interpretability: The system successfully compressed diverse high-performing rules into coherent "rule families" and generated qualitative explanations. For instance, a complex Boolean rule was translated into a narrative explaining that long call durations indicate engagement, while recent contacts suggest the customer is reachable.

Significance and Claims
The paper claims that BOOLXLLM successfully bridges the gap between formal, faithful explanations and human-understandable narratives. By combining the logical rigor of symbolic Boolean rules with the linguistic expressiveness and contextual awareness of LLMs, the framework offers an explainable AI system that is both theoretically grounded and accessible to non-experts. The work highlights the promise of combining symbolic reasoning with language-based models to advance human-centered XAI, demonstrating that such integration can improve interpretability and usability while maintaining competitive predictive performance.