The Joy of Dressing Is an Art: Outfit Generation Using Self-Attention Bi-LSTM

Abstract

Fashion represents one’s personality, what you wear is how you present yourself to the world. While in traditional brick & mortar stores, there is staff available to assist customers which results in increased sales, online stores rely on recommender systems. Proposing an outfit with-respect-to the desired product is one such type of recommendation. This paper describes an outfit generation framework that utilizes a deep-learning sequence classification based model. While most of the literature related to outfit generation is regarding model development, the segment describing training data generation is still not mature. We have proposed a novel approach to generate an accurate training dataset that uses the latent distance between positive and random outfits to classify negative outfits. Outfits are defined as a sequence of fashion items where each fashion item is represented by its respective embedding vector obtained from the Bayesian Personalised Ranking- Matrix Factorisation (BPR-MF) algorithm which takes user clickstream activity as an input. An outfit is classified as positive or negative depending on its Goodness Score predicted by a Bi-LSTM model. Further, we show that applying Self-Attention based Bi-LSTM model improved the performance (AUC), relevance (NDCG) by an average 13%, 16% respectively for all gender-categories. The proposed outfit generation framework is deployed on Myntra, a large-scale fashion e-commerce platform in India.