1: \begin{abstract}

2: This work proposes a rapid global solver for nonconvex low-rank matrix

3: factorization (MF) problems that we name \ours.

4: %by leveraging efficient convex lifting

5: %\red{steps, which we name \ours}.

6: Through convex lifting steps, our method efficiently escapes

7: saddle points and spurious local minima ubiquitous in noisy real-world

8: data, and is guaranteed to always converge to the global optima.

9: Moreover, the proposed approach adaptively adjusts the rank for the

10: factorization and provably identifies the optimal rank for MF

11: automatically in the course of optimization through tools of manifold

12: identification,

13: and thus it also spends significantly less time on parameter tuning

14: than existing MF methods, which require an exhaustive search for this

15: optimal rank.

16: On the other hand, when compared to methods for solving the lifted

17: convex form only, \ours leads to significantly faster convergence and

18: much shorter running time.

19: Experiments on real-world large-scale recommendation system problems

20: confirm that \ours can indeed effectively escapes spurious local

21: solutions at which existing MF approaches stuck, and is magnitudes faster

22: than state-of-the-art algorithms for the lifted convex form.

23: \end{abstract}

24: