Triple Factorization-Like Symmetric and Nonnegative Latent Factor Models for Undirected, Sparse and Large-Scaled Networks

Abstract

Non-negative latent factor (NLF) models have great capabilities of extracting useful knowledge from a symmetric, sparse and high-dimension (SHDiS) matrix with positive constraints. For the purpose of obtaining the NLFs, the tradition SNLF model is constructed based on the doublefactorization (DoF)-based matrix factorization (MF) technique, however it suffers from the low prediction accuracy. In order to improve the performance, an improved SNLF model in terms of triple factorization MF technique is proposed. Whereas, such a better prediction accuracy might be obtained at an evident cost of the time efficiency. Aiming at finding a fine balance between the model performance and the time cost, a novel triple factorization-like SNLF (TFL-SNLF) model is developed by introducing a pre-given non-negative symmetric matrix and a non-negative and symmetric LF matrix. Based on this established model, a single latent factor-dependent nonnegative additive gradient descent (AGD) update algorithm is designed for obtaining desired LFs. Experiments on two actual industrial data sets illustrate that the novel TFL-SNLF model can not only have a better performance of the prediction accuracy on missing data than the DoF-based SNLF model, but also is superior at the efficiency over the TrF-based SNLF. Hence, the proposed TFL-SNLF model is more applicable according to the industrial engineering.