Kernel Fashion Context Recommender System (KFCR): a Kernel Mapping Fashion Recommender System Algorithm using Contextual Information

Abstract

In Fashion, Recommender System represents a growing trend. They enable to offer to the customer online fully personalized shopping experience. Many known names on the Fashion market such as Asos (asos.com) or Zalando (zalando.com), has already bet on this technology to retain customers, leading to boosting profits. Multiple filtering approaches exist, developed in order to cope with all the inherent challenges driven by the lack of information and the ephemeral nature of fashion items. However, even if the methods adopted are very varied, the main motive remains the same: reducing at all costs, the margin of error in order to produce the nearest real prediction. Accuracy, scalability, flexibility, and performance have become the keywords when it comes to creating a fully skilled Recommender System. To achieve these objectives, it appears that including in the model user‟s context information, such as time, location, mood, occasion, weather, or people‟s influence can be the answer. Obviously, not including contextual information is eluding a fundamental element in the decision-making process for the purchase of a particular piece of clothing. In this paper, we decided to apply to the Fashion domain issues the scalable context-aware algorithm to better target the tastes of the customer producing predictions as close to his preferences as possible. This new algorithm named KFCR uses the kernel mapping framework developed by Ghazanfar et al. (KMR) complements with contextual information. We used the RentTheRunway Fashion dataset which indexes more than 100000 rented garments in the renttherunway.com website. We evaluated and compared this new system to the original KMR , as well as to other widely used context-aware approaches like post-filtering techniques. Once evaluated, the KFCR was found to be more accurate than both non-context-aware and context-aware approaches.