Improved Latent Factor Models for Undirected and Sparse Networks with Large Scales

Abstract

Nonnegative latent factor (NLF) models are highly efficient in extracting useful information from undirected, high-dimension, nonnegative and sparse (SHiDS) networks, which are commonly encountered in various industrial applications. The conventional double factorization (DF)-based matrix factorization (MF) technique contributes to obtaining LFs from SHiDs networks, whereas might lead to a fairly low prediction accuracy due to the limitation of the number of variables in NLF matrix. In order to tackle this problem, a novel NLF model via the trip factorization (TF)-based matrix MF technique is proposed, which is definitely equipped with 1) fulfilling the constrains on the symmetry; 2) a higher prediction accuracy; and 3) the nonnegativity of the NLF matrix. Experimental results on two data sets from real industrial applications demonstrate that the proposed TF-based NLF model has a better capability of guaranteeing the desirable performance at a bit expense of computational efficiency.