A Comprehensive Review of Deep Learning Integration in Recommender Systems: Taxonomy, Challenges, and Future Directions

The integration of deep learning (DL) into recommender systems (RS) has significantly reshaped how personalized content is generated and delivered across diverse domains. Traditional recommendations such as collaborative filtering and content-based filtering struggle to cope with the increasing complexity, diversity, and sparsity inherent in modern user-item data. DL techniques, however, can learn rich, non-linear mappings from multi-modal and large-scale data inputs. This is a comprehensive survey that synthesizes the outcome of ۴۰ peer-reviewed papers published in the time period ۲۰۲۳-۲۰۲۵ to provide a fine-level taxonomy of DL architectures like CNNs, RNNs, Transformers, GNNs, and Autoencoders with multimodal and hybrid architectures. We categorize and compare and contrast these models in terms of methodology, application area (e.g., healthcare, academia, streaming media, e-commerce), and key challenge areas like cold-start, scalability, interpretability, and fairness. Furthermore, this paper advocates for an integrated pipeline through AutoML, federated learning, and pretraining with contrast to overcome the barriers related to personalization, privacy, and versatility. Through state-of-the-art model benchmarking and future trends such as LLM-based personalization and ethics-aware design, this survey not only recapitulates latest progress but also charts the future direction to the next generation of trustworthy and intelligent recommender systems.

Keywords:

Deep Learning , Recommender Systems , Graph Neural Networks , AutoML , Contrastive Learning , Personalization , Fairness , Federated Learning , LLMs , Multi-Modal Fusion