Neural News Recommendation with Attentive Multi‑View Learning and Personalized Attention

Personalized news recommendation is essential for modern online news platforms to alleviate information overload, improve user experience, and increase engagement and revenue. The article outlines three core challenges: modeling user interests from reading histories, representing news content comprehensively, and ranking candidate news effectively.

To address the diversity of news information, the authors propose NAML (Neural Attentive Multi‑View Learning), presented at IJCAI 2019. NAML employs a news encoder that extracts representations from multiple views—title, body, category, and sub‑category—using word‑level and view‑level attention mechanisms, and a user encoder that aggregates high‑information news via news‑level attention. Click prediction is performed by the inner product of user and candidate news vectors.

Experimental evaluation on a month‑long MSN news dataset shows that NAML significantly outperforms traditional baselines, with ablation studies confirming the importance of each view and attention component. Visualizations reveal higher attention weights for body text and category information, and highlight salient words and news items influencing predictions.

Building on this, the authors introduce NPA (Neural Personalized Attention), presented at KDD 2019, which adds personalization to the attention mechanisms. A word‑level personalized attention uses a user‑specific query vector derived from the user ID embedding, while a news‑level personalized attention captures user preferences over news items. Negative sampling is employed to handle class imbalance, and the model is trained as a K+1‑way classification task.

Further experiments demonstrate that NPA consistently surpasses baselines and the non‑personalized variant, with significant gains in both accuracy and computational efficiency. Additional analyses show the critical role of negative sampling ratios and the combined effect of word‑ and news‑level personalized attention.

In conclusion, the two proposed methods—multi‑view learning for heterogeneous news representation and personalized attention for modeling user interest diversity—achieve state‑of‑the‑art performance in neural news recommendation, offering a practical solution for large‑scale, personalized news services.