Predicting Purchase Behavior using Visually Generated Product Gallery Networks

Modern e-commerce recommendation systems recommend users products through purchase prediction off of historical purchase data. This signal however has limitations as new and long tail products have little to no such signal to exploit. One signal however that influences user purchase behavior, especially in verticals such as fashion, is visual. In this paper we explore how visual similarity and object detection can together be used to predict fashion purchase behavior without using any purchase network based features. We formulate the problem as a network inference problem through creating a network consisting of product and gallery images with productproduct edges, derived using visual similarity, and product-gallery edges, derived using a combination of object detection and visual similarity. We evaluate our approach through triplets sampled from the Amazon purchase relationship.