# Neural News Recommendation with Negative Feedback

arXiv: https://arxiv.org/abs/2101.04328 · HTML: https://mlsanity.com/paper/2101.04328/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 12 Jan 2021.

## Built on by (most cited first, 5 of 5)

- [Modeling User Viewing Flow Using Large Language Models for Article Recommendation](https://mlsanity.com/paper/2311.07619/) (2023)
- [UltraWiki: Ultra-fine-grained Entity Set Expansion with Negative Seed Entities](https://mlsanity.com/paper/2403.04247/) (2024)
- [Quality-aware News Recommendation](https://mlsanity.com/paper/2202.13605/) (2022)
- [Time Matters: Enhancing Pre-trained News Recommendation Models with Robust User Dwell Time Injection](https://mlsanity.com/paper/2405.12486/) (2024)
- [Peeling Back the Layers: An In-Depth Evaluation of Encoder Architectures in Neural News Recommenders](https://mlsanity.com/paper/2410.01470/) (2024)

## Abstract (arXiv)

News recommendation is important for online news services. Precise user interest modeling is critical for personalized news recommendation. Existing news recommendation methods usually rely on the implicit feedback of users like news clicks to model user interest. However, news click may not necessarily reflect user interests because users may click a news due to the attraction of its title but feel disappointed at its content. The dwell time of news reading is an important clue for user interest modeling, since short reading dwell time usually indicates low and even negative interest. Thus, incorporating the negative feedback inferred from the dwell time of news reading can improve the quality of user modeling. In this paper, we propose a neural news recommendation approach which can incorporate the implicit negative user feedback. We propose to distinguish positive and negative news clicks according to their reading dwell time, and respectively learn user representations from positive and negative news clicks via a combination of Transformer and additive attention network. In addition, we propose to compute a positive click score and a negative click score based on the relevance between candidate news representations and the user representations learned from the positive and negative news clicks. The final click score is a combination of positive and negative click scores. Besides, we propose an interactive news modeling method to consider the relatedness between title and body in news modeling. Extensive experiments on real-world dataset validate that our approach can achieve more accurate user interest modeling for news recommendation.