A Convex Approach to Low Rank Matrix Approximation with Missing Data