HHAGP RecSys: A system to model personalized size-fit fashion recommendations
✨A Holistic Hybrid Adaptive Geometric Deep Learning and Perceptual (HHAGP) Fashion Recommender System with Applied Metaverse for size-fit problems in online shopping✨ by Darius Singh.
The paper was presented at the 9th International Conference on Business Analytics and Intelligence (ICBAI) organized by IIM Bangalore (15 - 17 December 2022). It was also presented at the Paper Presentation Competition by the Department of Mathematics, Sri Venkateswara College (University of Delhi) where it won the best presentation award (5 April 2023).
Abstract
Clothing is the most frequently returned item on e-commerce platforms. Most of the garments are returned due to a poor fit or poor appearance on the customer’s body. Returning purchased items often not only has an impact on the carbon footprint but also leads to an unsatisfied customer experience. In India alone, the return rate for purchases online is 25-40% and the key contributor is size-fit. Our system not only makes recommendations based on the user’s style preferences but also takes into account how the garment fits and looks on the user’s body ensuring a more personalized fashion recommendation for the user. Previous research on fashion recommendation systems has focused on modelling recommendations based on user history, item-item compatibility, body shape, or styling preferences using social media or user history patterns. We introduce a Holistic Hybrid Adaptive Geometric and Perceptual (HHAGP) Recommender System for individualized fashion recommendations. Our proposed system addresses the lack of cross-brand standardisation of clothing sizes and the lack of a fashion recommendation methodology that takes into account both users’ physical characteristics and style preferences. The HHAGP recommender system complements the existing class of content-based and collaborative fashion recommender systems by evaluating the recommendations geometrically and perceptually in order to further subset them so that they are highly individualized and user-specific. To achieve this goal, the recommendations obtained from the implicit content-based and collaborative recommender system are passed through a Monocular-to-3D Virtual Try-On Network (M3D-VTON) to visualize the recommendation on the user’s body and then filtered by a Visual Geometry Group model (VGG) that assigns a perception score to the recommendation. The measurements for the different sizes of the filtered recommendations are geometrically matched to the user’s physical characteristics in order to obtain the best-fit recommendation size. An explicit feedback loop continuously adapts the system to the user’s preferences. The results demonstrate a robust user-centric system that provides user-specific fashion recommendations considering both qualitative and quantitative characteristics. Our system combines traditional user history-based collaborative recommendation systems with user-clothing fit and user-clothing perception combined with a hybrid feedback loop to model accurate clothing recommendations. The augmented reality-based meta-virtual surrounding helps the consumer to visualise different environments with the recommended apparel. This paper presents a novel solution that provides user preference-based fashion recommendations that are perceptually appealing and complement the user’s body shape and size. This approach when applied will result not only in reducing the returns of the apparel but also help in developing an immersive dataset for Indian body fits by size.
For more information
- Presentation of the paper by the author: