Usage of deep neural networks as recommendation engines. Review of selected approaches

Abstract

This paper reviews selected neural network architectures used in the context of e-commerce recommendation engines. Subsequent approaches are presented in order from the most straightforward applications to the latest advancements. It compares the models in terms of generalisation capabilities as well as the possibility to utilize external features. The formalization of the essential learning procedures is presented, as well as the derivation of equivalence between autoencoder neural networks and the matrix factorization approach. In conclusion - modern deep learning architectures composed of the autoencoder component and dense low-dimensional features (called encodings) have much stronger predictive power and can replicate, without the loss of generality, the behavior of matrix-factorization approaches.