Active Adaptation of Expert-Based Suggestions in Ladieswear Recommender System LookBooksClub via Reinforcement Learning

摘要

Fashion recommendation is one of the developing fields in e-commerce. Many different types of recommender systems exist with their own advantages and disadvantages. In this paper we create a recommender system for ladieswear that utilizes all recommender system approaches: collaborative filtering, content-based, demographic-based and knowledge-based. Using stylists' suggestions, we created distance space for items, user clusters and connected item features to users' characteristics. Stylist initial ratings were used to solve the cold-start problem. We adopted the Upper Conditional Bounds (UCB) algorithm for active selection of items which should be suggested. The system was designed with strong constraints dictated by the business process. The system worked for one month and estimated with 64 % of "likes" received for its suggestions, while the well-known Rocket Retail system shows only 55 % of "likes" after five years of its use.