Lightweight structure-aware explanations for rule-based models: application to radiotherapy risk communication in...

Rule-based models are often considered inherently interpretable because their explicit logical structure allows users to trace predictions back to human-readable rules. However, in complex domains such as healthcare, multiple rules involving numerous attributes may be activated simultaneously, making it difficult for end users to identify the most influential factors behind a prediction. This can limit practical transparency, particularly in clinical shared decision-making, where clear and trustworthy explanations are essential. An example is communicating the risk of side effects following radiotherapy for breast cancer, where understanding the determinants of estimated risk may support informed treatment choices.

We propose a lightweight, structure-aware attribution method for rule-based models inspired by information retrieval techniques. The approach represents the activated rules associated with a prediction as a document-like structure and applies a TF-IDF-inspired weighting scheme to estimate the relevance of individual input attributes. Unlike perturbation-based methods such as SHAP and LIME, the proposed approach operates directly on the internal model structure, without requiring input perturbations, surrogate models, repeated evaluations, or access to training data. This preserves the explanatory signal encoded in the rules while reducing computational complexity and execution time. The method is evaluated through a technical comparison with SHAP and LIME and a user-centered study involving potential end users.

Experimental results show moderate agreement between the TF-IDF-inspired ranking and established attribution methods, identifying attributes that are behaviorally informative with respect to the implemented predictor. Faithfulness analysis shows that direct rule frequency achieves higher perturbation faithfulness, highlighting a trade-off introduced by the proposed weighting scheme. Complementarily, the user study indicates that the complete explanation pipeline, combining relevance weighting, semantic aggregation, and user-oriented presentation, is perceived as more interpretable and useful than raw rule-based explanations. These findings suggest that information-retrieval-inspired weighting can provide a lightweight mechanism for organizing rule evidence, while semantic abstraction and user-oriented presentation can facilitate its communication to end users.