UTAM: A Unified Thematic Mapping Approach for Cross-Domain Personalization and Cold Start Alleviation

Abstract

Cross-domain recommendation systems are essential for delivering personalized experiences across diverse content areas. However, aligning user preferences across distinct domains, especially in cold-start scenarios where users have no interaction history in the target domain remains a significant challenge. To bridge this gap, we propose UTAM (Unified Thematic Alignment and Mapping), a novel framework that leverages thematic clustering of user preferences in a source domain to enable recommendations in a target domain. By capturing semantically aligned user-topic engagement patterns, UTAM effectively transfers user intent across domains. We evaluate UTAM on Amazon Movies and Books datasets using ten different experimental configurations that explore various dimensionality reductions, clustering thresholds, and engagement metrics. In cold-start scenarios, UTAM outperforms its single-domain variants. Specifically, UTAM achieves a precision of 0.770, recall of 0.854, and F1-score of 0.783. In comparison, the single-domain Movies model records precision 0.7005, recall 0.2506, F1-score 0.3174, while the single-domain Books model achieves precision 0.6309, recall 0.3142, F1-score 0.3649. These results highlight UTAM’s robustness and effectiveness in alleviating the cold-start problem while delivering accurate, scalable, and interpretable cross-domain recommendations.