Optimizing Seed Set for New User Cold Start

He-Da Wang, Ji Wu

Abstract: Users newly enter a recommender system can not get personalized recommendation due to the lack of personal profiles. An interview process that asks new users to rate a set of items (the seed set) will help user modeling and improve user experience. Traditional seed set generation approaches often concentrate on item-wise properties instead of aiming at finding the optimal seed set. We propose a simple random optimization technique to search for the optimal seed set, which considers the seed set as a whole and performs a random search by reducing the prediction error on validation set. By off-line experiments on the Movie Lens 10M data set, we show that the proposed approach performs as well as the state-of-the-art method called Greedy Extend, and the proposed approach needs significantly less computational cost to reach the same prediction error as the best baseline on validation set.