Handling Cold-Start Collaborative Filtering with Reinforcement Learning