A Fashion Recommendation System Based on Transfer Learning and Deep Ensemble Model: Enhancing Style Prediction

Abstract

The fashion industry has rapidly grown, driven by dynamic trends and consumer preferences. In this competitive environment, personalized recommendation systems play a crucial role in enhancing customer experience and driving sales. This paper introduces an innovative fashion recommendation system that applies transfer learning and an ensemble of deep learning models to improve style prediction accuracy. The system utilizes five pre-trained models VGG16, VGG19, MobileNet, DenseNet, and Xception each contributing feature extraction from fashion images. The approach includes key stages like preprocessing, feature extraction, and classification. Preprocessing involves resizing, normalization, and noise reduction, ensuring high-quality inputs for the models. Feature extraction draws significant fashion attributes, such as texture, colour, and pattern, from the convolutional layers of these models. By combining the outputs of the five models, the ensemble classifier boosts prediction accuracy and robustness. The model was tested on a comprehensive fashion dataset with 44,000 images across 144 categories, achieving an impressive 96.5% accuracy, 96.4% precision, 96.8% recall, and an F1-score of 96.5%. Comparative analysis demonstrates that the proposed ensemble approach outperforms traditional machine learning models and current state-of-the-art deep learning techniques. These results emphasize the system's potential to deliver highly accurate, personalized fashion recommendations, making it an effective solution for e-commerce platforms.