Effects of preprocessing and training biases in latent factor models for recommender systems

Abstract

Abstract Latent factor (LF)-based models are highly efficient in addressing high-dimensional and sparse (HiDS) matrices raised in big-data-related applications like recommender systems. While linear biases have proven to be effective in improving the prediction accuracy and computational efficiency of LF models, their individual and combinational effects in such performance gain remain unclear. To address this issue, this work aims at studying the effects of linear biases in LF models for recommender systems. Based on careful investigations into existing methods, we categorize frequently adopted biases into two classes, i.e., preprocessing bias (PB) and training bias (TB). Subsequently, we deduce the training objectives and parameter updating rules of LF models with different PB and TB combinations. Experimental results on three HiDS matrices generated by real recommender systems show that (a) each PB/TB does have positive/negative effects in the performance of an LF model; (b) Such effects are partially data dependent, however, some specific PB/TB can bring stable performance gain to an LF model; and (c) several PB and TB combinations appear significantly more effective in improving an LF model's performance when compared to their peers.