Hyper-parameter-evolutionary latent factor analysis for high-dimensional and sparse data from recommender systems

摘要

High-dimensional and Sparse (HiDS) data generated by recommender systems (RSs) contain rich knowledge regarding users' potential preferences. A Latent factor analysis (LFA) model enables efficient extraction of essential features from such data. However, an LFA model relies heavily on its hyper-parameters like learning rate and regularization coefficient, which must be chosen with care. However, traditional grid-search-based manual tuning is extremely time-consuming and computationally expensive. To address this issue, this study proposes a hyper-parameter-evolutionary latent factor analysis (HLFA) model. Its main idea is to build a swarm by taking the hyper-parameters of every single LFA-based model as particles, and then apply particle swarm optimization (PSO) to make its both hyper-parameters, i.e., the learning rate and regularization coefficient, self-adaptive according to a pre-defined fitness function. Experimental results on six HiDS matrices from real RSs indicate that an HLFA model outperforms several state-of-the-art LF models in terms of computational efficiency, and most importantly, without loss of prediction accuracy for missing data of an HiDS matrix. (c) 2020 Elsevier B.V. All rights reserved.