Attention-Driven Fusion of Pre-Trained Model Features for Superior Recommender Systems

Recommender systems are essential in online platforms and e-commerce as they help users find personalized and relevant products. This study presents a novel two-stage fashion product recommendation system designed to significantly enhance recommendation relevance through the integration of attention-based deep learning architectures. In the first stage, a Dual Attention Network is employed to categorize query product images by leveraging deep feature representations extracted from pre-trained Xception and VGG16 models. These features are refined through attention modules and fused to predict the product category. The second stage involves recommending relevant products for the query image based on similarity measures. The proposed two-stage approach achieves superior performance, with 92.63% accuracy on the Fashion Product Images dataset and 81.06% accuracy on the Shoe dataset, surpassing existing models. This work’s main contributions consist of the development of a Dual Attention Network specifically designed for product categorization, the effective fusion of attention-based features from multiple pre-trained models, and the successful application of these techniques to improve fashion product recommendations.