Visual similarity-based fashion recommendation system

Abstract Fashion recommendation systems are designed to retrieve a ranked list of catalog images similar to the target catalog image. These systems are crucial in helping customers move toward potential buying decisions from recommended products. Catching true correlation between similar product recommendations and user satisfaction will exponentially improve sale experience of companies. One of the major tasks in fashion recommendation systems is accurately finding similar images for a given product image. In this chapter, we outline a visual similarity-based fashion recommendation system that can be used in e-commerce especially for shoe and handbag recommendations. The system consists of several modules including a generative adversarial network (GAN) module for creating image representations (feature vectors) and an effective vector search library that acts as a database for querying similar images. We provide comparisons for our GAN-based recommendation results with pretrained convolutional neural networks (CNNs) on new catalog shoe and handbag images. Major contributions of this chapter can be summarized as follows: by utilizing GANs we are able to develop a robust and reliable fashion recommendation system, visual similarity-based recommendations are employed to create similar product search for e-commerce applications, and we also provide detailed comparisons for extracting high-quality, accurate, and efficient vector feature representation of product images using deep neural networks for fashion domain.