A Tertiary Study on Approaches for Developing Recommender Systems

Abstract

Recommender systems (RS) have emerged as essential tools in the digital age, mitigating information overload through personalized recommendations. This tertiary study conducts a systematic mapping of scientific production on RS over the last five years, focusing on the main approaches used, the contexts in which they are applied, and the ongoing challenges. A review of 58 secondary studies identified eight predominant approaches: Collaborative Filtering, Content-Based Filtering, Deep Learning, Graph Neural Networks, Knowledge Graph Embeddings, Reinforcement Learning, Large Language Models, and Hybrid Methods. The results reveal that, despite the increasing adoption of advanced techniques such as GNNs and LLMs, traditional methods, including KNN, SVD, and MF, remain widely used. Furthermore, a lack of systematic comparative studies evaluating the effectiveness of techniques in different contexts was observed. This work contributes to the identification of gaps and to guiding future studies on recommender systems, and presents a taxonomy of the identified approaches, highlighting their respective strengths and limitations.