Combining Knowledge Graphs with Personalized News Recommendation Systems

The rapid shift of news consumption from print to digital has created information overload, prompting the development of personalized news recommendation systems that can filter and rank articles for individual users.

Two main approaches exist: traditional feature‑engineering methods that match TF‑IDF vectors of users and articles, and deep‑learning methods that learn implicit semantic representations for matching in latent space.

Integrating knowledge graphs into recommendation brings richer semantic relations, enabling deeper user interest discovery, more meaningful explanation paths, and higher user acceptance. A simple selection algorithm extracts news‑relevant entities from large corpora, expands them by one or two hops, and assigns weighted importance to retain only the most relevant triples.

The system introduces explicit and implicit topic nodes, connects them with entities, and enriches the graph with collaborative relations, improving both recommendation accuracy and interpretability.

Experiments on Microsoft News and Meituan datasets compare three graph variants (generic KG, news‑relevant KG, and KG with collaborative/topic enhancements) using baseline models (LDA+DSSM, BERT) and demonstrate consistent performance gains across personalization and classification tasks.

A three‑layer model—entity representation (using KGAT), context embedding, and information distillation—injects knowledge‑enhanced vectors into news representations. Multi‑task learning and reinforcement‑learning‑based anchor knowledge graphs further boost accuracy and provide interpretable recommendation paths.

Extensive ablation studies confirm the contribution of each module, and visualizations illustrate how important entities receive higher weights and how anchor graphs generate diverse explanation paths.

The presentation concludes with a Q&A covering event graphs, product inventory graphs, pattern mining, open‑source knowledge graphs, and practical advice for building domain‑specific graphs, emphasizing the relevance of graph learning for real‑world recommendation systems.