Reinforcement Learning with Knowledge Graphs for Personalized News Recommendation

¹ Department of Information Technology, Trinty College of Engineerng & Research, Pune, India
² Department of Electronics and Communication Engineering, St.Peter’s Institute of Higher Education and Research, Chennai, India
³ Department of Electronics and Communication Engineering, Nitte Meenakshi Institute of Technology, Nitte (Deemed to be University), Bengaluru, India
⁴ Independent researcher

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

In recent years, the rapid growth of online media and personalized news recommendations have become essential for reducing information overload and improving user engagement. However, traditional models significantly rely on semantic similarity and suffered to capture dynamic user preferences. Moreover, the extensive dissemination of misinformation poses serious limitations, necessitating recommendation systems that are both adaptive and knowledge aware. To overcome these limitations, a Reinforcement Learning with Knowledge Graphs (RL-KG) framework for personalized news recommendations. Initially, data was gathered from the MIND dataset, which consist of millions of users’ news interactions enriched with titles, abstracts, and entity annotations. During preprocessing, text normalization and entity linking are employed for semantic clarity and structured knowledge representation. Furthermore, feature extraction integrates Bidirectional Encoder Representations from Transformer (BERT) embeddings for contextual semantics with Knowledge Graph embeddings for entity-level reasoning. Moreover, the Reinforcement Learning (RL) model was utilized for sequential decision-making procedures in which user histories form states, candidate news represents actions, and rewards provide engagement. Finally, the recommendation module integrates hybrid features and RL policies to deliver personalized, diverse news suggestions. The experimental results demonstrate that RL-KG achieves higher accuracy (96.77%), precision (95.55%), recall (96.45%), and F1-score (93.76%) than the existing model Recommendation for Mitigation (Rec4Mit).