Personalized News Recommendation System Based on Knowledge Graphs

News consumption has shifted from print to digital, creating information overload; personalized news recommendation systems are needed to improve user experience.

Two main approaches exist: traditional feature‑engineering methods using TF‑IDF matching, and deep‑learning methods that learn implicit semantic representations of users and articles.

Incorporating a knowledge graph adds rich semantic relations, enabling deeper user interest discovery and providing explainable recommendation paths that increase user satisfaction.

The news knowledge graph is built by extracting seed entities from a large news corpus, expanding one or two hops, weighting relations, and adding explicit topic nodes (e.g., categories) and implicit topic nodes learned via LDA.

The proposed model consists of three layers: an entity representation layer using KGAT to aggregate neighbor information, a context embedding layer that captures entity position, frequency, and type, and an information‑distillation layer that weights entity importance and fuses it with the original news vector; the framework supports multi‑task learning.

Experiments on Microsoft News and Meituan datasets compare a generic KG, the constructed news KG, and baselines, showing consistent improvements in personalized recommendation and news classification tasks, with ablation studies confirming the contribution of each module.

For explainability, an anchor knowledge graph is generated per article via a reinforcement‑learning MDP, enabling multiple interpretable paths between articles; quantitative and visual analyses demonstrate higher quality and quantity of explanations compared to baselines.

The presentation concludes with a Q&A covering event graphs, product inventory graphs, pattern mining, open‑source KG resources, and the balance between using public versus proprietary knowledge graphs.