Transformers em sistemas de recomendação: um estudo comparativo com modelos de machine learning

Abstract

Recommender Systems (RS) are essential for personalizing the user experience on digital platforms, but they face challenges such as cold-start and data sparsity. This work performs a comparative analysis of the ranking capability between a model based on the Transformers architecture and established Machine Learning (ML) algorithms — SVD++, ItemKNN, and a Popularity-based model — for movie recommendation. Using the MovieLens datasets (100K and 1M), the study evaluates the effectiveness of the models in generating relevant rankings, especially in critical scenarios such as coldstart. The evaluation metrics focus on the quality of the ranking, such as Normalized Discounted Cumulative Gain (NDCG), Precision@k, and Recall@k, complemented by error metrics (Root Mean Squared Error (RMSE), Mean Absolute Error (MAE)) of the scores that generate these rankings. The methodology employs k-fold cross-validation (k=10). The results indicate that the Transformers model, although presenting higher recall in some scenarios, was consistently outperformed by traditional models in ranking metrics, such as NDCG and Precision@k. However, they presented greater robustness in cold-start situations for items with few interactions. It is concluded that, although it does not yet surpass consolidated methods in all metrics, the Transformers shows potential to be explored in specific recommendation contexts, contributing to the literature with insights on the applicability and performance of Deep Learning models.