Weightless Neural Network WiSARD Applied to Online Recommender Systems

Abstract

Recommender systems generally are made to predict user preferences' for items. However, in high dimensional datasets this task demands high computational costs. Taking into account that data distribution changes through time, it is important that online recommender systems have a fast retraining process in order to keep the model updated, delivering accurate predictions. Therefore, we propose a new approach for recommender systems using a weightless neural network, denominated WiSARD. We show that our proposal increases training and prediction processing speed, without decreasing the quality of predictions. First results show that our proposal is 306% faster than the improved regularized singular value decomposition (IRSVD), a well-known state-of-the-art algorithm. Moreover, our proposal still had an improvement of 3.7% regarding the mean absolute error (MAE). We show how to apply the WiSARD algorithm for online recommender systems, its drawbacks, and insights for further research.