A Fashion recommendation system using an autoencoder network with integrated feature extraction by CNN models

Abstract

Due to the rapid rise in living standards, the fashion and textile industries have experienced remarkable growth in the fast fashion industry in recent years. On e-commerce websites, the users can find numerous types of fashion items and it can consume lot of time and effort for them to find exactly what they want. Due to this, there is a need for an effective recommendation system that allows users to easily search, sort, and receive pertinent information. In the present, most of the available recommendation systems provide recommendations for a user based on his or her previous activities and behaviors. The problem with this approach is that user preferences can change overtime and new trends can come and go very frequently. Therefore, we cannot expect that the user will like the same type of fashion items even for a short period of time because in the fashion industry designs move very quickly. The other popular approach is that suggesting items based on the information available on similar users. This is not successful either because the given user’s preferences are not considered. This paper presents an automated fashion recommendation system that produce recommendations based on an image input by the user. This research also provides more insights on applicability of convolutional neural networks and autoencoder networks for feature extraction along with a comparison of the performance of each of the selected models. This work also suggests a hybrid approach for achieving better performance and evaluation of different ensemble techniques used for combining the outputs of individual models is also provided. The final outcome demonstrates how a hybrid strategy is helping to get results that are more accurate than individual models and selection of the most suitable ensemble technique is also important to further increase the accuracy of the overall system.