Deconvolving Feedback Loops in Recommender Systems