An investigation of pre-training approaches for matrix factorization-based recommender systems.

Abstract

Recommender Systems (RSs) consist of a field of research and application with the goal of retrieving relevant items for a user. Since the open Netflix Prize challenge for performance improvement in RSs, they have constantly been built by representing users and items as latent factors, more commonly known as embeddings, which are often randomly initialized and updated during the training stages. When looking at the greater Machine Learning (ML) area, different areas of application obtained performance improvement through Transfer Learning, such as the boost obtained in the Computer Vision (CV) tasks after the proposal of models like VGG or AlexNet, or the one achieved in Natural Language Processing (NLP) tasks, especially after the popularization of Large Language Models (LLMs) such as the BERT and, more recently, the GPT model families. Unlike other application areas, however, Transfer Learning for RSs is not trivial since users and items are the entities, while in CV and NLP, the entities are images and words, respectively. This research aims to study possible applications of Transfer Learning for RSs, evaluating how unsupervised, self-supervised, and supervised embedding initialization impact the predictive performance of the models.