An Uncertainty-Aware Hybrid Floriculture Recommendation System Driven by YOLO Architectures

Floriculture recommendation systems often face difficulties caused by ambiguous visual cues, high morphological variation among species, and limited user preference data. This study proposes a hybrid framework that integrates deep visual recognition with uncertainty-aware decision logic to enhance recommendation reliability in both typical and cold-start conditions. The system combines a YOLO-based visual encoder, collaborative filtering, and a content-based similarity module. To address prediction ambiguity, two uncertainty measures are introduced: a Class Uncertainty Index (CDI), derived from classlevel F1 performance, and an Epistemic Uncertainty (EU) score based on feature-space relevance. These are incorporated into a Weighted Utility Score (WUS) that penalizes recommendations with high uncertainty. Experiments on a 112-species dataset show clear improvements in accuracy and F1-score compared with YOLO and Content-Based Filtering (CBF) baselines, especially for morphologically variable species such as Hydrangea macrophylla and Ping Pong. Additional evaluations demonstrate reduced error rates and effective cold-start handling.