Explainable Uncertainty-Aware Recommendation via Prefix-Tuned Natural Language Generation

Author : Sanjib Kumar Swain, Santosh Kumar Swain

Abstract : Explainable recommendation systems generate uniformly assertive rationales regardless of upstream uncertainty, eroding user trust when predictions are uncertain. We propose XU-Rec (eXplainable Uncertainty-aware Recommendation), a parameter-efficient architecture that maps uncertainty triples (ue, ua, Δ) to natural language hedging via a 198K-parameter prefix adapter conditioning frozen GPT-2 (117M parameters). The modality-agnostic design accepts uncertainty from any upstream source with architectural non-circularity enforcement (zero text dependency), validated through eight independence tests. We introduce HA-Rich (H × R × C), a multi-level metric measuring uncertainty-linguistic alignment across lexical, syntactic, and semantic dimensions. Empirical validation on collaborative filtering uncertainty (Amazon Movie Reviews, 100K reviews, 5 seeds) demonstrates learned hedging behavior: HARich = 0.233 }0.030 with strong correlation (ρ=0.65) between interaction-based uncertainty and linguistic hedging, using 650× fewer trainable parameters (198K vs. 128.6M) than full finetuning baselines. The architecture enables future deployment with multimodal perceptual uncertainty (video/audio) without modification.

Keywords : Explainable recommendation, uncertainty quantification, natural language generation, prefix tuning, hedging appropriateness.

Conference Name : International Conference on AI and Machine Learning in Life Science Engineering (ICAIMLLSE - 26)

Conference Place : Chennai, India

Conference Date : 19th Sep 2026