Damped Gauss-Newton algorithm for nonnegative Tucker decomposition

Algorithms based on alternating optimization for nonnegative Tucker decompositions (NTD) such as ALS, multiplicative least squares, HALS have been confirmed effective and efficient. However, those algorithms often converge very slowly. To this end, we propose a novel algorithm for NTD using the Levenberg-Marquardt technique with fast computation method to construct the approximate Hessian and gradient without building up the large-scale Jacobian. The proposed algorithm has been verified to overwhelmingly outperform “state-of-the-art” NTD algorithms for difficult benchmarks, and application of face clustering.