A HYBRID DEEP NEURAL ARCHITECTURE FOR PERSONALIZED SKINCARE RECOMMENDATION

Abstract

Personalized skincare recommendations require accounting for both user preferences and the biological suitability of products. This study investigates personalized skincare recommendations through a neural model that integrates user behavior and ingredient-level product information. A Multi-Head Attention Two-Tower neural architecture was developed that combined user latent embeddings with product embeddings derived from ingredient metadata. The model was trained using ranking-based objectives on the Amazon Beauty dataset with 5-core filtering, comprising 5,269 user–item interactions. Performance was evaluated using standard ranking metrics, including Hit Rate (HR@K), Recall@K, and nDCG@K. The model achieved a Hit Rate of 12.41% at top-10 recommendations, outperforming random selection by an order of magnitude. Stratified analysis by skin type revealed the highest effectiveness for sensitive skin (HR@10 = 21.9%) and the lowest for dry skin (HR@10 = 3.6%), which demonstrates the system’s ability to encode biologically relevant compatibility patterns. Rank distribution analysis confirmed that recommendations were not dominated by popularity, with relevant items consistently prioritized across users. The results confirmed that incorporating attention mechanisms with a dual-tower architecture enables biologically informed and personalized recommendations. The findings support integrating explicit skin profiles and ingredient-level embeddings to enhance safety, relevance, and user trust in skincare recommendation systems.