Human-in-the-Loop Framework for eXplainable Recommender Systems Based on Explanation Information Quality 10 views

Recently, eXplainable Recommender Systems (XRS) have emerged as a critical technology for enhancing system transparency and user trust by providing explicit justifications for recommendations. However, current evaluations focus on recommendation performance or subjective user study. Consequently, they fail to assess the information quality of generated explanations from a human perspective. In particular, the phenomenon of machine hallucination, where models generate plausible but factually incorrect explanations, severely undermines user trust. Despite this, there is a notable scarcity of optimization frameworks designed to explicitly mitigate misinformation while accurately reflecting actual user preferences. To overcome these limitations, this dissertation introduces a human-in-the-loop framework for XRS that evaluates the information quality of explanations and continuously improves the output based on this assessment. First, it proposes Veracity, a novel multidimensional evaluation metric grounded in Signal Detection Theory (SDT) that quantifies the continuous, subjective spectrum of human perception. This metric provides a diagnostic tool by independently measuring Fidelity (how accurately the explanation reflects the actual characteristics of an item) and Attunement (how well the explanation aligns with the user's specific tastes). Second, leveraging the proposed Veracity metric, this research develops HIVE (Human-In-the-Loop framework for explainable recommender systems on VEracity). HIVE directly updates item and user embeddings based on veracity-driven human feedback, thereby capturing shifting user preferences over time and correcting misinformation to improve recommendation quality. Third, to address the exploratory limitations of HIVE and mitigate the filter bubble effect, this work proposes the extended HIVE+ framework. HIVE+ integrates a Collective Novelty Alignment mechanism that dynamically explores neighborhoods within the embedding space. This allows the system to effectively discover latent user preferences for novel items without compromising factual correctness. Furthermore, to address the persistent cold-start problem inherent in recommender systems (RS), this dissertation presents a practical methodology utilizing Large Language Model-based personas to synthesize and augment preference data from a minimal initial sample. The effectiveness of the proposed metrics and frameworks was validated through offline simulations, LLM-based pseudo-user studies, and actual human subject experiments. The results demonstrate that HIVE+ achieves significant gains over state-of-the-art baselines in both recommendation performance and explanation veracity. Notably, the human subject experiments revealed that real users apply a significantly stricter and more conservative bias against machine-generated misinformation compared to simulated agents. This empirical finding underscores the critical necessity of ensuring high veracity in real-world RS. Ultimately, by presenting an integrated, human-centric approach to evaluating and optimizing the information quality of XRS, this dissertation establishes a robust foundation for designing next-generation, trustworthy RS aimed at holistic human understanding.

PHD (Doctor of Philosophy)

English

2026-04-21