류성은/황재성/마수드’s paper has been accepted in

Title: POPCORN: Popularity-decoupled Interest Matching for Personalized News Recommendation,
Author: Seongeun Ryu, Jaesung Hwang, Masoud Reyhani Hamedani, and Sang-Wook Kim
Abstract
To alleviate the information overload problem on online news plat-forms, personalized news recommendation aims to provide news articles aligned with users’ interests. Most conventional news recommendation methods suffer from popularity bias, in which popular news is recommended to users regardless of their genuine interests. To mitigate this, popularity-aware news recommendation methods regard popularity-driven clicks as bias in recommendation process. Although these methods show higher accuracy than the conven-tional ones, they still face the following three challenges inherited from conventional methods: (C1) discriminating click types on pop-ular news, (C2) topic-unaware popularity modeling, and (C3) dealing with interest signal attenuation. To jointly address these challenges, we propose POPCORN as a novel framework for POPularity-deCOupled inteRest matching in personalized News recommenda- tion, which consists of three key strategies: (1) Popularity-decoupled news modeling to prevent popularity from interfering with content representation and enable the distinction between interest-driven and popularity-driven clicks, (2) Candidate-guided user modeling to amplify a user’s interest signal for a candidate news based on its topic, and (3) Popularity-penalized interest matching to suppress rec-ommendations for popular news that is not genuinely interesting to the user. POPCORN is designed as a model-agnostic framework, and can be plugged into any news recommendation method without modifying its architecture. The results of our extensive experiments with three real-world datasets and eight state-of-the-art methods demonstrate that the methods augmented with POPCORN consis- tently outperform their original versions. All source code will be publicly available on our GitHub page upon acceptance of the paper.