Upbit Presents Research Paper on Personalized News Recommendation at International Conference on Information Retrieval

Dunamu, the operator of Upbit, announced on the 16th that a research paper by its machine learning (ML) team on personalized news recommendations was accepted at the international information retrieval conference SIGIR 2025 and was presented at the main conference.

This year, the main conference of SIGIR is being held from July 13th to 18th at Centro Congressi, located in Padua, Italy. Chungwon Park, a researcher on Dunamu's machine learning team, presented Dunamu's research results on its personalized news recommendation system at the conference on the 14th.

The paper is titled 'User Simulator Based on Large Language Model (LLM): A Methodology for Training News Recommendation Models Without Real User Interaction.' The paper describes an approach that uses virtual users generated by LLM, instead of real user data, to recommend news.

Specifically, Dunamu introduced the 'LAUS (LLM As User Simulator)' framework in the paper, which can alleviate dependence on users in existing news recommendation methods. According to Dunamu, traditional news recommendation required collecting user data such as click logs and news preferences for model training, which raised concerns about large-scale privacy breaches.

LAUS generates training data by creating virtual users instead of using real user data. As a result of this study, LAUS showed higher performance and lower latency than the traditional 'zero-shot' news recommendation method. In comparative evaluations of news recommendation systems in multiple languages, including English, the model trained with LAUS achieved performance similar to that of models trained with actual user data.

Researcher Park stated, "The quality of personalized news recommendation systems is directly linked to how accurately they provide users with the information they want, which is a key factor in improving service satisfaction. This research lays the foundation for providing more sophisticated recommendation services while enhancing customer information protection and operational efficiency."