Accuracy meets Diversity in a News Recommender System

Proceedings of the 29th International Conference on Computational Linguistics, pages 3778–3787 October 12–17, 2022. 3778 Abstract News recommender systems face certain challenges. These challenges arise due to evolving users’ preferences over dynamically created news articles. Diversity is necessary for a news recommender system to expose users to a variety of information. We propose a deep neural network based on a two-tower architecture that learns news representation through a news item tower and users’ representations through a query tower. We introduce diversity in the proposed architecture by considering a category loss function that aligns items’ representation of uneven news categories. Experimental results on two news datasets reveal that our proposed architecture is more effective compared to the state-of-the-art methods and achieves a balance between accuracy and diversity. 1 Introduction Reading the news has never been more common in people’s daily lives than it is now. Big names like Yahoo!, Google, and CNN have launched online news portals that users can access from anywhere to browse various news categories and find up-to- date information. However, finding the right content is a challenge. With so much information available online, selecting relevant news has become a time-consuming and challenging task. A news recommender system (NRS) offers solutions to the information overload problem and provides relevant and interesting news recommendations to users (Raza and Ding, 2021b). In the state-of-the-art of NRS, the news stories that users have read in the past are used to infer their interests and preferences. However, there is usually frequent content updating in a news domain. We show the news reading behaviour of a typical user in Figure 1 as an example. Figure 1: A user’s behaviour during news reading We see in Figure 1 that this user reads about the ‘COVID-19 cases in Ontario’ during time t1, and then read about ‘Canadian children 5 to 11 could become eligible for COVID-19 vaccine…” during t2, and then his following action (read/click) is reading the news ‘COVID-19 travel ban on tourists’ during t3. We also observe that there is a transition in user preferences to other topics, for example, the elections, and politics in the later time. This shows that a user in an NRS generally changes his/her interests over time. Some of the user interests are long-term which shows a habit or personality, while some of the user interests are short-term, which may erupt due to some trending news, a sudden interest in an event and so. It is important to consider the accuracy as well as the diversity of news items while providing recommendations to users in an NRS (Raza and Ding, 2021a). This is a motivation for this research. A common practice in the recommender systems is to design a two-tower architecture (Wang et al., 2019), where, first, a retrieval model retrieves a subset of related items from a large corpus in response to a user’s query, and, then a ranking model ranks the retrieved items based on users’ actions (clicks or ratings). The quality of retrieved items plays …

Content truncated to 500 words