Is This News Still Interesting to You?: Lifetime-aware Interest Matching for News Recommendation Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conferen

Seongeun Ryu, Yunyong Ko, Sang-Wook Kim · 2025

Read Is This News Still Interesting to You?: Lifetime-aware Interest Matching for News Recommendation Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conferen by Seongeun Ryu, Yunyong Ko, Sang-Wook Kim (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-13064.txt.

Introduction

Seongeun Ryu, Yunyong Ko, and Sang-Wook Kim present in 2025 a news recommendation method that minds the clock. A reader's past clicks age, and each story has its own shelf life, yet most systems match interests as if time stood still. Their frame, LIME, builds time into the match in three ways.

First, it scores the age of each clicked story for that reader and topic. Next, it weighs the reader's profile with the candidate story's timing in view. Last, a freshness step favors stories still in their prime. On two real data sets the method beat a wide range of rivals, and its parts lifted other models when added.

Catalog record

CREATORSeongeun Ryu, Yunyong Ko, Sang-Wook Kim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-13064

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Is This News Still Interesting to You?: Lifetime-aware Interest Matching for News Recommendation Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM ’25), November 10–14, 2025, Seoul, Republic of Korea DOI: 10.1145/3746252.3761047 ISBN: 979-8-4007-2040-6/2025/11 CCS: Information systems Recommender systems CCS: Information systems Personalization CCS: Computing methodologies Neural networks Seongeun Ryu, Yunyong Ko, Sang-Wook Kim Abstract . Personalized news recommendation aims to deliver news articles aligned with users’ interests, serving as a key solution to alleviate the problem of information overload on online news platforms. While prior work has improved interest matching through refined representations of news and users, the following time-related challenges remain underexplored: ( C1 ) leveraging the age of clicked news to infer users’ interest persistence , and ( C2 ) modeling the varying lifetime of news across topics and users . To jointly address these challenges, we propose a novel L ifetime-aware I nterest M atching framework for n E ws recommendation, named LIME , which incorporates three key strategies: (1) User-Topic lifetime-aware age representation to capture the relative age of news with respect to a user-topic pair, (2) Candidate-aware lifetime attention for generating temporally aligned user representation, and (3) Freshness-guided interest refinement for prioritizing valid candidate news at prediction time. Extensive experiments on two real-world datasets demonstrate that LIME consistently outperforms a wide range of state-of-the-art news recommendation methods, and its model-agnostic strategies significantly improve recommendation accuracy. † † cc-license: by 1. Introduction Personalized news recommendation aims to provide news articles that align with the interest of an individual user, which serves as a pivotal role for alleviating information overload on online news platforms. A general approach in existing works ( Wu et al. 2019c ; Wu et al. 2019d ; Wu et al. 2019b ; An et al. 2019 ; Wu et al. 2019a ; Wang et al. 2020 ; Liu et al. 2020b ; Qi et al. 2021c ; Mao et al. 2021 ; Mao et al. 2022 ; Chen et al. 2023 ; Yang et al. 2023 ; Li et al. 2022 ; Ko et al. 2025 ) typically follows a three-step pipeline: (1) ( news representation ) a news article is represented based on its attributes (e.g., title, body, topic) as a news embedding, (2) ( user representation ) the embeddings of historically clicked news of a user are aggregated into a user embedding, and (3) ( interest matching ) the degree of matching between the user and candidate news embeddings is estimated as an interest score. Accordingly, prior works on personalized recommendations have focused on enhancing news modeling ( Wu et al. 2019c ; Wu et al. 2019d ; Wu et al. 2019b ; Wu et al. 2019a ; Wang et al. 2020 ; Kim et al. 2022 ; Liu et al. 2020b ; Mao et al. 2021 ; Mao et al. 2022 ; Yang et al. 2023 ; Ko et al. 2025 ) , user modeling ( Wu et al. 2019c ; Wu et al. 2019d ; An et al. 2019 ; Mao et al. 2021 ; Mao et al. 2022 ; Ko et al. 2025 ) , or interest matching ( Qi et al. 2021c ; Chen et al. 2023 ; Qi et al. 2021b ; Qi et al. 2021a ; Li et al. 2022 ) . Figure 1 . Age, lifetime, and freshness of news. Figure 2 . Topic-wise and user-wise news lifetimes. Time-sensitive nature of the news domain. Compared to other recommendation domains, the news domain is inherently more time-sensitive ( Wu et al. 2020a ; Liu et al. 2020b ; Wu et al. 2023 ; Qi et al. 2021b ; Chen et al. 2023 ; Bae et al. 2023 ) . For instance, approximately 80% of news articles in the MIND dataset ( Wu et al. 2020a ) received their last click within 48 hours of publication ( Wu et al. 2020a ) . In contrast, items in the Netflix dataset 1 1 1 https://www.kaggle.com/datasets/netflix-inc/netflix-prize-data continued to receive interactions over several months, with 80% of total interactions spread over an average of 32 months ( Bae et al. 2023 ) . This phenomenon indicates that the news content (e.g., title, body) quickly loses its validity to users over time , which distinguishes the news domain from others. To address this time-sensitive nature of the news domain, recent studies ( Liu et al. 2020b ; Qi et al. 2021b ; Chen et al. 2023 ; Bae et al. 2023 ; Cho et al. 2021 ) have attempted to incorporate temporal attributes, such as age , freshness , and lifetime of news articles (see Figure 1 ), into news recommendation. For example, HyperNews ( Liu et al. 2020b ) aims to capture the temporal dynamics of a news article by incorporating its age (i.e., elapsed time from publish time) information into the news content embedding. LANCER ( Bae et al. 2023 ) introduces the concept of news lifetime , defined as the elapsed time from publish time to the point at which a predefined top- m\% (e.g., m =90) of clicks have occurred. Challenges . Although existing works considering the time-sensitive nature of news have achieved higher accuracy, the following challenges still remain under-explored: (C1) Leveraging age of news in click history. Intuitively, a user’s click implies that the clicked news is valid to her at that time. This indicates that the age of clicked news in a user’s click history contributes to better understanding her interest in a temporal perspective . Existing methods ( Liu et al. 2020b ; Qi et al. 2021b ; Chen et al. 2023 ; Bae et al. 2023 ) , however, have focused only on the age of candidate news in order to filter out outdated news from recommendation, rather than user modeling, which limits the potential to capture users’ interests accurately in a temporal aspect. (C2) Varying lifetime across topics and users. It is known that news articles have shorter lifetime (e.g., 36 or 48 hours) than those of items in other recommendation domains ( Bae et al. 2023 ) . Looking more closely, the lifetime of news content may vary across topics . For example, sport news about game results is quickly consumed and expires within a few hours, whereas travel news about landmarks or tips tends to remain valid for several days or even weeks (See Figure 2 (a)). Moreover, even for the same topic, users may differ in durations in which they remain interested. For instance, a user A may consume preferable news articles quickly, while a user B may consume them for a longer period (See Figure 2 (b)). Although one existing work ( Bae et al. 2023 ) considers lifetime of news articles, it just adopts a single fixed lifetime for all news, failing to exploit its full potential in news recommendation. In summary, existing methods have the following limitations: (1) they may recommend news articles that align with a user’s interest but are no longer valid at that time , and (2) they fail to capture a user’s temporal aspect of

Read the complete text