Recommendation Fashion Products Using Collaborative Filtering and Variational Autoencoders (CF-VAE)

The growth of the digital fashion industry requires accurate and personalized recommender systems for users. Previous studies in the area of fashion recommender system able to understand user preferences. However, limited user interactions with items, as well as the presence of new users or items without historical data, still remain a problem. The prior works that utilize content-based and visual-based approach, only use type of information and are less effective in providing relevant recommendations when data is limited. To address this problem, this study proposes hybrid Collaborative Filtering Variational Autoencoders (CF-VAE) approaches as a solution that combines strength CF in capturing user interaction patterns and the ability of VAE to form latent representations of user preferences to generate more accurate and personalized fashion product recommendations. The CF-VAE model is implemented employing an encoder and decoder architecture combined with Matrix Factorization (MF) as a baseline model. The experiment was conducted using a fashion rental dataset with 11,032 interactions of 5,438 users and 2,685 fashion items. The evaluation results show that CF-VAE outperforms both baseline MF and pure Multinomial-VAE (Mult-VAE). CF-VAE achieves 223.4% improvement in Precision@10, CF embeddings to Mult-VAE contribute an additional 52% improvement, demonstrating the effectiveness of the hybrid architecture. Therefore, as a hybrid solution, this study is proven to be capable and effective in handling sparse data with a matrix density of 99.92% and producing more relevant recommendations for users.