KnowEntityRec: entity-centric knowledge perception graph neural network for news recommendation

Accurate news recommendations are essential in today’s digital environment. However, due to the timeliness of news, there is a great demand for recognising the latest related entities in the emerging news. This paper introduces KnowEntityRec, which leverages the Knowledge Perception Module (KPM) to create dynamic, real-time cross-dataset knowledge graphs, enabling the detection of co-occurring and relevant entities and thereby improving recommendation accuracy. The KPM integrates cross-knowledge-graph information and constrains multi-hop expansion through a hop limitation mechanism, enabling controllable and semantically diverse entity augmentation. Extensive experiments conducted on the MIND-small and MIND-large datasets demonstrate the effectiveness of the proposed method. KnowEntityRec achieves a 0.63% improvement in AUC over the state-of-the-art PNR-LLM on MIND-large, and the improvement is statistically significant under a paired t-test. Consistent gains are also observed across multiple evaluation metrics on MIND-small. These results indicate that KPM provides a lightweight yet effective approach for knowledge-enhanced news recommendation.