Author(s) :

Sherry Chen.

Volume/Issue :

Abstract :

Skin type match is a key consideration of fit in beauty product recommendation, yet most deployed recommendation systems do not have users’ skin type information readily available, and standard precision-based metrics may not capture skin compatibility. This paper proposes a methodology to infer customers’ skin type and integrates this signal into the hybrid beauty product recommender. Using the Sephora dataset (8,494 products, 294,722 reviews), purpose-built skin-type signals were built— including a skin-type-aware collaborative filtering matrix and a trained skin-type classifier—into various hybrid skincare recommendation configurations. Eight domain-specific metrics are proposed to evaluate the recommenders. A Bayesian weight optimization procedure using a Tree-structured Parzen Estimator (TPE) was applied to find optimal weights for the hybrid system. The results yield four key findings. First, personalization lifts are statistically significant across all four tested skin-type profiles on skin compatibility and routine coherence. Second, no single configuration dominates on all metrics: combining product content with skin profile achieves the highest skin-type compatibility and rank sensitive precision among profile-aware variants; combining collaborative filtering with skin profile leads on diversity and serendipity; full hybrid provides balanced performance across all metrics. Third, profile weighting produces genuine inter-profile differentiation confirmed by positive adjusted personalization scores across all hybrid variants. Fourth, Bayesian optimization identifies the skin-type classifier as the dominant signal and reveals that enforcing a minimum content weight improves out-of-sample generalization. These results confirm that these inferred skin-type signals and the use of a multi-criteria evaluation framework can significantly improve the quality of the beauty product recommendation.