Recommendations from Sparse Comparison Data: Provably Fast Convergence for Nonconvex Matrix Factorization