Balancing preference and compatibility: A multiobjective optimization framework for outfit recommendation

Garment recommendation in fashion e-commerce requires a practical balance between personalized user preference and outfit compatibility. Existing recommendation methods usually combine these two goals into a single scalar objective, which limits their ability to represent diverse trade-offs in apparel selection. This study proposes a multiobjective outfit recommendation (MOOR) framework that formulates outfit recommendation as a multiobjective optimization problem and searches for Pareto-optimal outfit solutions. To estimate the two objectives, the framework incorporates a heterogeneous graph-based user preference model for capturing sparse and higher-order preference signals, and a multimodal outfit compatibility model for assessing visual, textual, and attribute-level coherence between garments. A task-specific evolutionary search strategy is further introduced to explore candidate top–bottom combinations while preserving recommendation diversity. Experiments on the IQON3000 dataset show that the proposed framework provides strong preference and compatibility estimation and generates diverse trade-off solutions with favorable outfit-level quality under the reported evaluation protocol. These findings support the value of explicitly modeling the preference–compatibility trade-off in garment recommendation systems.