Adaptive multi-scenario matching through uncertainty-aware heterogeneous user modeling.

Multi-scenario matching is pivotal for retrieving relevant items across diverse contexts. However, existing methods often neglect the fine-grained user-specific interaction patterns essential for personalized recommendation. Furthermore, these approaches implicitly assume a similar level of reliability in behavioral evidence across users, thereby overlooking the severe disparity arising from interaction sparsity. Consequently, learned representations can become unreliable for long-tail users. To address these challenges, we propose PER-UHM, a unified Uncertainty-Aware Heterogeneous User Modeling Framework. We first build on PERSCEN as the base architecture, which employs a user-specific feature graph and vector quantization to extract fine-grained, scenario-aware preferences. We then generalize this architecture into PER-UHM by incorporating an Adaptive Uncertainty-aware Prior Fusion (AUPF) mechanism on the user tower. Unlike standard approaches, this module introduces a Dual-Codebook Decoupling strategy with Uncertainty Gating to balance profile priors and behavioral evidence according to estimated evidence reliability. Extensive experiments on the KuaiRand-Pure and Alimama datasets show that PERSCEN and PER-UHM achieve strong and consistent performance for both active and long-tail users.