Personalized Outfit Recommendation using Style Vector Extraction

Abstract

In many online fashion stores, outfit recommendations are based on item compatibility, offering the same outfits to all users regardless of individual preferences. To address this limitation, personalized recommendations based on users’ past item clicks are crucial. It is important to extract a style because the style characteristics of items clicked by the user in the fashion domain influence the recommendation. Given the ambiguous nature of style, we propose using deep learning to extract style vectors by applying ViT encoders in metric learning. This approach trains distances between input data pairs within the same class to be close and those between different classes to be far, utilizing a similarity matrix to ensure similar labels aren’t treated as negative in case of different classes. After extracting styles, we compute vector similarities between items in user clickstreams and items in recommended out- fits, proposing a model that sorts recommendations by highest similarity. Comparing our proposed model with current recommendation systems, random sorting, and similarity matrix-based methods, our model showed superior performance, albeit slightly lower than the currently recommended model likely due to UI interference. To validate this, we conducted a user study revealing that our model performed best, indicating that providing similarity information is more effective than relying on non-personalized recommendation systems.