Abstract In the era of information overload, sequential recommendations are crucial for capturing users’ dynamic preferences; however, they suffer from challenges such as data sparsity and noise stemming from superficial interactions. Previous approaches to address these issues have evolved from recurrent neural networks (RNNs) and attention mechanisms to graph neural networks, yet they still struggle to learn precise and robust item representations. This study aims to bridge this gap by proposing a novel and unified framework. In the proposed method, knowledge graph embeddings are refined and optimized through contrastive predictive coding. Subsequently, these enriched representations are integrated with users’ dynamic preferences, extracted by a sequential model based on LSTM and self-attention, to generate highly personalized and accurate recommendations. The key finding of our research, based on experiments conducted on three benchmark datasets, demonstrates the significant superiority of the proposed model, achieving an AUC of 0.97 on the MovieLens-1M dataset.
Khaligh et al. (Mon,) studied this question.