A hybrid explainability framework for recommender systems using SHAP with natural language interpretations

Abstract

Recommender systems based on deep neural networks are widely recognized for their high predictive accuracy, but their opaque nature limits transparency and interpretability, crucial qualities needed to ensure accountability, fairness, and trust, especially as the right to explanation is increasingly regarded as a legal and ethical obligation in high-stakes decision-making. This study presents a hybrid explainability framework for recommender systems that combines SHAP (Shapley Additive Explanations) with natural language justifications to enhance the explainability of recommender system outputs. The framework integrates both local and global SHAP-based explanations. Local explanations analyze instance-level Shapley values to identify features that positively influence individual recommendations, while global Shapley values, computed at the item level, capture the dominant features characterizing the recommended item, which are then translated into item-level textual descriptions. These two explanation layers are then used to generate coherent, human-readable justifications for recommended items using large language models. The approach is implemented in a career recommendation engine based on aptitude profiles from the Occupational Information Network (O*NET). Expert evaluations on a five-point Likert scale yielded mean scores between 3.7 and 4.4 across various user profiles, indicating moderately high acceptance of the explanations generated by the framework. The findings suggest that the framework enhances transparency, fairness, accountability, and user trust in the recommendation process.