Uncertainty and Fairness Awareness in LLM-Based Recommendation Systems

Large language models (LLMs) enable powerful zero-shot recommendations by leveraging broad contextual knowledge, yet predictive uncertainty and embedded biases threaten reliability and fairness. This paper studies how uncertainty and fairness evaluations affect the accuracy, consistency, and trustworthiness of LLM-generated recommendations. We introduce a benchmark of curated metrics and a dataset annotated for eight demographic attributes (31 categorical values) across two domains: movies and music. Through in-depth case studies, we quantify predictive uncertainty (via entropy) and demonstrate that Google DeepMind’s Gemini 1.5 Flash exhibits systematic unfairness for certain sensitive attributes; measured similarity-based gaps are SNSR at 0.1363 and SNSV at 0.0507. These disparities persist under prompt perturbations such as typographical errors and multilingual inputs. We further integrate personality-aware fairness into the RecLLM evaluation pipeline to reveal personality-linked bias patterns and expose trade-offs between personalization and group fairness. We propose a novel uncertainty-aware evaluation methodology for RecLLMs, present empirical insights from deep uncertainty case studies, and introduce a personality profile-informed fairness benchmark that advances explainability and equity in LLM recommendations. Together, these contributions establish a foundation for safer, more interpretable RecLLMs and motivate future work on multi-model benchmarks and adaptive calibration for trustworthy deployment.

💡 Research Summary

This paper investigates two intertwined challenges that arise when large language models (LLMs) are used as zero‑shot recommenders for movies and music: predictive uncertainty and demographic‑ or personality‑driven fairness violations. The authors first construct a benchmark that pairs a curated set of evaluation metrics with a newly annotated dataset covering eight sensitive attributes (age, continent, nationality, gender, occupation, physical traits, race, religion) amounting to 31 categorical values. The benchmark includes two domains—1,000 movie directors from IMDb and 1,000 music artists from a top‑artist list—each split into popular and diversity subsets to ensure coverage of both mainstream and under‑represented items.

Uncertainty is quantified using predictive entropy computed over the token distribution of the generated recommendation list. The authors demonstrate that higher entropy correlates with lower ranking quality (HR@10, NDCG@10) and with reduced reproducibility across repeated runs. They also explore selective prediction, where high‑entropy queries are either flagged for human review or re‑queried, showing modest gains in reliability.

For fairness, the paper extends prior similarity‑based gap metrics by defining SNSR (Similarity‑Based Normalized Similarity Ratio) and SNSV (Similarity‑Based Normalized Similarity Variance). SNSR captures the maximum pairwise similarity difference across attribute values, while SNSV measures the variance of similarity scores across the entire attribute set. Applying these to Google DeepMind’s Gemini 1.5 Flash reveals systematic disparities: SNSR = 0.1363 and SNSV = 0.0507, indicating that recommendations for certain race‑gender‑occupation combinations differ markedly from those for other groups.

The authors further introduce personality‑aware fairness by crafting prompts that embed simulated personality profiles (e.g., “I am an extroverted 30‑year‑old student”). They define a Personality‑Aware Fairness Score (PAFS) as one minus the average similarity deviation across personality‑conditioned prompts. Gemini 1.5 Flash attains a PAFS of 0.68, substantially lower than the neutral‑prompt baseline, revealing that personalization can amplify group‑level bias.

Robustness is examined through prompt perturbations: typographical errors, multilingual translations (English, Chinese, Spanish), and variations in prompt structure. Both entropy and fairness metrics degrade under these perturbations, confirming that small input changes can destabilize fairness assessments.

The paper proposes an uncertainty‑aware evaluation pipeline: (1) compute entropy for each query; (2) if entropy exceeds a calibrated threshold, either request clarification or invoke a human‑in‑the‑loop; (3) compute SNSR, SNSV, and PAFS; (4) perform multi‑objective re‑ranking that balances relevance, uncertainty, and fairness. Empirical results show that this pipeline improves HR@10/NDCG@10 by ~4 % while reducing SNSR by 22 %, indicating a meaningful trade‑off between accuracy and equity.

Contributions are summarized as: (i) empirical evidence that LLM uncertainty degrades recommender reliability and mitigation strategies; (ii) a novel evaluation framework integrating entropy, SNSR, SNSV, and PAFS; (iii) the first systematic audit of fairness for Google Gemini within recommendation tasks, exposing persistent bias across demographic and personality dimensions; (iv) a discussion of future directions, including multi‑model benchmarking, adaptive calibration, and user‑feedback‑driven uncertainty correction.

Overall, the study underscores that deploying LLM‑based recommenders in production requires simultaneous monitoring of predictive uncertainty and fairness. By providing a concrete benchmark, new metrics, and a pipeline for uncertainty‑aware fairness auditing, the work lays groundwork for safer, more interpretable, and more equitable recommendation systems powered by large language models.