Abstract: Scientific paper recommendation systems serve as an effective solution to the problem of information overload in academic databases. This study proposes a knowledge-graph-based method employing attention embeddings for the task of scientific paper recommendation to enhance the effectiveness of recommendations. This method initially constructs a collaborative knowledge graph to integrate user behavior with paper attribute information and optimizes node vector representations using the TransR approach. Subsequently, it introduces an attention sequence module that employs an attention propagation mechanism to learn node features and utilizes a sequence attention mechanism to capture the temporal preferences of users from their reading sequences. Finally, the model calculates match scores between researchers and candidate papers to generate personalized recommendation lists. Experiments conducted on a dataset provided by the "Blockchain Laboratory" have validated the effectiveness of the model. Experimental results indicate that the proposed model significantly improves recommendation recall rates, capturing the dynamic interests of researchers more accurately. This study not only enhances the performance of scientific paper recommendation systems but also provides new perspectives and tools for understanding and predicting the evolution of researcher interests.

Key words: Scientific papers recommendation, Dynamic preferences, Knowledge graph, Attention embedding, Self-attention mechanisms