Estudo comparativo entre sistemas de recomendação tradicionais e baseados em deep learning: análise de desempenho e eficiência computacional

Abstract

Recommender systems are digital tools used on online platforms to personalize user experiences and facilitate the discovery of relevant items. This work presents a comparative study between traditional and deep learning-based approaches for movie recommendation, using The Movies Dataset. Two traditional models – FunkSVD and TF-IDF – and two deep learning-based models – AutoRec and DistilRoBERTa with triplet loss fine-tuning – were evaluated. The evaluation considered predictive metrics (Precision@10, Recall@10, and NDCG@10) and computational efficiency (training time, inference time, and peak RAM usage). FunkSVD achieved the best performance in Precision@10 (0.7365) and Recall@10 (0.4871), outperforming both AutoRec and DistilRoBERTa in these aspects. Although DistilRoBERTa reached the highest NDCG@10 (0.8336), it also had the longest training time (22.17 minutes). TF-IDF showed the lowest memory consumption but had the worst predictive results. The findings indicate that, despite the potential of deep learning-based models, traditional approaches like FunkSVD can offer competitive performance with lower computational cost. In certain contexts, simpler methods may even outperform more complex solutions, especially when data is scarce or computational resources are limited.