Dress for Success: An Incremental Approach Towards Fashion Compatibility Featurization

Abstract

Fashion recommendation systems have grown in popularity in recent years, fueled by the rise of e-commerce and the demand for personalized and convenient shopping experiences. However, predicting fashion compatibility remains a complex task due to the subjective nature of fashion styles and trends. In this work, we propose an incremental approach to fashion compatibility featurization, which leverages careful feature extraction methods to improve compatibility prediction performance. Our method achieves comparable results to current state-of-the-art approaches on common fill-in-the-blank and outfit compatibility tasks, despite using a comparatively simpler methodology. Additionally, our approach successfully captures both similarity properties between clothing items and compatibility relationships in building outfits, making it a robust and versatile approach for fashion compatibility prediction across popular compatibility datasets. We conduct extensive experiments on Maryland Polyvore, Polyvore Outfits, and Polyvore-D to demonstrate the effectiveness of our proposed method across these popular datasets.