Document Type : Original/Review Paper

Authors

• Mahdie Norozian ¹
• Razieh Rastgoo ²

¹ Electrical and Computer Engineering Faculty, Semnan University, Semnan, Iran

² Electrical and Computer Engineering Department, Semnan University, Semnan, Iran

Abstract

The fashion industry is continuously evolving, with new styles emerging at an unprecedented rate. In this dynamic landscape, fashion recommender systems play an important role in providing personalized suggestions based on users’ preferences and individual styles. A key challenge in this domain is predicting the compatibility of fashion items, which is essential for building effective outfit recommendation systems. In this study, we propose a novel Transformer-based model for outfit item compatibility prediction that jointly exploits visual information from fashion images and semantic information from textual descriptions. To improve robustness and generalization, the proposed model employs a composite loss function that combines Focal Loss with L1 and L2 regularization, helping to address class imbalance while reducing overfitting. Experimental evaluation on the Polyvore Outfits dataset demonstrates the effectiveness of the proposed approach, achieving an AUC of 0.972 and an FITB accuracy of 91.5%. These results demonstrate the potential of the proposed model for improving the accuracy and personalization of fashion recommendation systems.

Keywords

• Recommender Systems
• Fashion
• Compatibility Prediction
• Deep Learning
• Transformers

Main Subjects

• H.6.5.2. Computer vision