A Study on Developing a Personalized Makeup Recommendation Model Using Deep Neural Networks Based on Korean Skin Physiological Indicators: With Cross-National Data Comparison

The purpose of this study was to develop and validate a deep neural network (DNN)-based personalized makeup recommendation system using skin physiological indicators from the Skin Genomics Information Platform (SGIP). The dataset comprises 3,148 data points collected over 11 years (2011–2021) from 18 countries, with Korean data (192 participants: 96 males, 96 females) selected as the primary analysis subject due to high data completeness (96.0%), balanced gender and age distribution, and consistent measurement protocols across all nine indicators. The study employed descriptive statistics, correlation analyses, DNN performance evaluation, and ARIMA/Prophet time series forecasting. Results showed significant gender and age differences in skin parameters. The DNN model achieved superior performance with 89.2% accuracy, 0.882 F1-score, and 2.34 RMSE, outperforming Random Forest (84.7%) and SVM (82.3%). Feature importance analysis identified skin hydration (18.5%), brightness (16.2%), and pore size (14.8%) as critical factors influencing makeup recommendations. ARIMA forecasting predicted 2025-2034 trends showing increases in brightness and hydration, with decreases in sebum and elasticity, providing quantitative insights for mid-to-long-term product development strategies. The multilayer DNN effectively learned complex nonlinear relationships among skin indicators, while Integrated Gradients improved interpretability. This mixed-method approach with multinational data overcame small-scale study limitations, enabling comprehensive skin analysis. The model supports evidence-based personalized recommendations and diagnostic device development. Long-term forecasts provide essential data for whitening, moisturizing, and anti-aging product strategies.