[Paper Review] Neural News Recommendation with Attentive Multi-View Learning

Chuhan Wu, Fangzhao Wu+4 more|arXiv (Cornell University)|Jul 12, 2019

Topic Modeling19 references33 citations

TL;DR

This paper introduces NAML, a neural news recommender that learns unified news representations from multiple views (title, body, and category) with word-, news-, and view-level attentions, and builds user representations from browsed news for improved recommendation.

ABSTRACT

Personalized news recommendation is very important for online news platforms to help users find interested news and improve user experience. News and user representation learning is critical for news recommendation. Existing news recommendation methods usually learn these representations based on single news information, e.g., title, which may be insufficient. In this paper we propose a neural news recommendation approach which can learn informative representations of users and news by exploiting different kinds of news information. The core of our approach is a news encoder and a user encoder. In the news encoder we propose an attentive multi-view learning model to learn unified news representations from titles, bodies and topic categories by regarding them as different views of news. In addition, we apply both word-level and view-level attention mechanism to news encoder to select important words and views for learning informative news representations. In the user encoder we learn the representations of users based on their browsed news and apply attention mechanism to select informative news for user representation learning. Extensive experiments on a real-world dataset show our approach can effectively improve the performance of news recommendation.

Motivation & Objective

• Motivate improved news and user representations by leveraging multiple information sources (title, body, category) beyond a single news view.
• Develop an attentive multi-view news encoder to fuse diverse views with word- and view-level attention.
• Construct a user encoder that selects informative browsed news to form user representations.
• Implement a click predictor based on inner product between user and candidate news representations.
• Evaluate on real-world MSN News data and demonstrate performance gains over baselines.

Proposed method

• Propose a news encoder with four components: title encoder with CNN and word-level attention; body encoder with CNN and word-level attention; category encoder transforming category IDs through dense layers; and a view-level attention to weight the importance of title, body, category, and subcategory views.
• Compute a unified news representation by combining views using view-level attention: r = alpha_c r^c + alpha_sc r^sc + alpha_t r^t + alpha_b r^b.
• In the user encoder, learn user representation by applying a news-level attention over browsed news representations: u = sum_i alpha_i^n r_i.
• Use an inner product y_hat = u^T r_c as the click probability predictor, with negative sampling during training to form a pseudo K+1 classification task.
• Train end-to-end using negative log-likelihood with softmax normalization over positive and sampled negatives (K=4).
• Experiment setup: MSN News dataset with 10k users, 42,255 news items, 489,644 positives, 6,651,940 negatives; embeddings 300/100, CNN filters 400, window 3, dropout 20%, Adam optimizer, batch 100.

Experimental results

Research questions

• RQ1 Can multi-view (title, body, category) information improve news representation learning over single-view approaches?
• RQ2 Do word-level and view-level attentions improve the selection of informative words and views for news representation?
• RQ3 Can a attention-based user encoder leveraging browsed news improve user representations for click prediction?
• RQ4 How does NAML perform against baseline neural and traditional recommender methods on real-world data?

Key findings

• NAML outperforms all baselines on AUC, MRR, nDCG@5, and nDCG@10; the best variant (NAML*) achieves significant improvements (p<0.001) over all baselines.
• Using the body view yields stronger performance than title or category alone, and combining all views yields further gains.
• Word-level, news-level, and view-level attentions each contribute to performance, with their combination providing the best results.
• View-level attention reveals body view generally more informative, yet category_view can be highly informative for topics, validating the multi-view approach.