Personalized online search for fashion products

Abstract

In this paper we develop a methodology for personalizing online search for fashion products. Most search functions on fashion retail sites currently rely on objective data, such as colors or brands, to filter their products. Adding a subjective component of fashion style would allow for a more personalized and relevant search experience for the end user. The proposed methodology is based upon a topic modeling technique – latent Dirichlet allocation – which has been successfully used for classifying unstructured text data. This technique is used to quantitatively define distinct fashion styles based upon text obtained from clothing product information available through APIs with affiliate networks. Using the fashion style definitions, individual clothing brands and looks are then classified. We compare the performance of the proposed methodology with Genostyle's proprietary methodology (Genostyle is a fashion styling analytics company). An experiment was executed to display custom recommendations made using each methodology to participants. The team measured performance by comparing median 5-point Likert scale responses to the recommended looks. Results indicate the latent Dirichlet allocation methodology has higher median Likert responses for the top recommended style. However, the Likert scores for Genostyle’s methodology, the latent Dirichlet allocation methodology, and the control group are statistically indistinguishable (p=0.05). Given these results, further experimentation with more diverse participation or presentation of looks may give insights into how to better fashion style predictions that more closely match consumers’ preferences.