AN EXTENDED FRANK-WOLFE METHOD WITH "IN-FACE" DIRECTIONS, AND ITS APPLICATION TO LOW-RANK MATRIX COMPLETION.

Motivated principally by the low-rank matrix completion problem, we present an extension of the Frank-Wolfe method that is designed to ind uce near-optimal solutions on lowdimensional faces of the feasible region. This is accomplished by a new approach to generating "in-face" directions at each iteration, as well as through new choice rules for selecting between inface and "regular" Frank-Wolfe steps. Our framework for generating in-face directions generalizes the notion of away steps introduced by Wolfe. In particular, the in-face directions always keep the next iterate within the minimal face containing the current iterate. We present computational guarantees for the new method that trade off efficiency in computing near-optimal solutions with upper bounds on the dimension of minimal faces of iterates. We apply the new method to the matrix completion problem, where low-dimensional faces correspond to low-rank matrices. We present computational results that demonstrate the effectiveness of our methodological a pproach at producing nearly optimal solutions of very low rank. On both artificial and real datasets, we demonstrate significant speedups in computing very low rank nearly optimal solutions as compared to the Frank-Wolfe method (as well as several of its significant variants). [ABSTRACT FROM AUTHOR]