Evolution of Neural Collaborative Filtering for Recommender Systems

Abstract

Recommender Systems are a highly active area in research and development that has taken advantage of the recent progress in artificial intelligence and deep learning algorithms. Collaborative Filtering approaches have utilized neural networks for modelling complex nonlinear relationships regarding the interactions of users and items, with numerous commercial platforms utilizing such systems for providing personalized recommendations to their users. In this work, we present the evolution of the field and the most influential approaches, from simpler neural network models expanding matrix factorization techniques, to increasingly more complex architectures. We report notable applications and highlight key differences between research and production settings. At the same time, we note that the evaluation approaches followed by the literature vary, and we underline the significance of testing models during production with methods such as A/B tests and the measurement of key performance indicators, aside from offline testing.