Active Inference as an Intrinsically Explainable Recommender Architecture via Epistemic-Pragmatic Score Decomposition

Abstract

Explainability and user control remain major challenges in modern recommender systems, especially in neural and sequential ranking models where recommendation scores are often produced from opaque latent representations. Existing post hoc explanation methods may describe model outputs, but they often fail to reveal the actual decision criteria used to rank items. This limitation becomes more important as recommender systems must balance relevance, novelty, diversity, uncertainty, and popularity risk. This study proposes an intrinsically explainable recommender architecture grounded in Active Inference. The framework treats recommendation as a transparent decision process in which each item is evaluated through interpretable components related to preference alignment, information gain, uncertainty, and exposure risk. By making these components part of the scoring process itself, the approach provides explanations by design rather than after the ranking has already been produced. The proposed architecture also supports user control by allowing recommendation behavior to adapt to changing preference beliefs and exploration needs. Evaluation against standard recommender baselines shows that the approach can maintain competitive ranking performance while improving transparency and control beyond accuracy objectives. These findings position Active Inference as a promising foundation for explainable, adaptive, and auditable recommender systems.

Keywords

Active Inference · recommender systems · explainability · epistemic value · pragmatic value · expected free energy · uncertainty-aware ranking

Suggested Citation

Chokhawala, B., & Mohammad, A. F. (2026). Active Inference as an Intrinsically Explainable Recommender Architecture via Epistemic-Pragmatic Score Decomposition. 44th International RAIS Conference. https://doi.org/10.5281/zenodo.22802029

BibTeX

@inproceedings{chokhawala2026activeinference,
  author    = {Chokhawala, Bhagyeshkumar and Mohammad, Atif Farid},
  title     = {Active Inference as an Intrinsically Explainable Recommender Architecture via Epistemic-Pragmatic Score Decomposition},
  booktitle = {44th International RAIS Conference},
  year      = {2026},
  doi       = {10.5281/zenodo.22802029},
  url       = {https://rais.education/wp-content/uploads/0689.pdf}
}