- This research challenges the conventional understanding of inverse probability weighting (IPS) and newly identifies the cause of performance improvements in recommended models.

The University of Seoul announced that a paper by Professor Jong-June Jeon’s research team, titled "Revisiting IPS in Recommendation Models: Unveiling Its Impact on Model Performance," has been accepted for presentation at WSDM (Web Search and Data Mining) 2026, one of the world's most prestigious international conferences in data mining and machine learning.

WSDM, organized by the Association for Computing Machinery (ACM), is one of the leading international conferences in web search, recommendation, and data mining. It presents the latest research in search and data mining across the web and social web. Recognized as a top-tier conference by both the National Research Foundation of Korea and the Korean Institute of Information Science and Technology, it receives thousands of submissions each year and maintains a highly competitive acceptance rate of about 15–20%.

▶ The Selection Bias Problem in Recommender Systems and the Proposed Methodology

The accepted paper offers a new perspective on the effectiveness of IPS, which is widely used to address selection bias in recommender systems. The research team showed theoretically that, when the model is correctly specified, the risk function with inverse propensity weights and the conventional risk function without these weights can produce the same optimal solution. This finding challenges the common belief that the performance of existing inverse propensity weight-based methods results from unbiased risk minimization.

This result suggests that the causal methodology (inverse propensity weighting) introduced to separate advertising exposure effects from individual preferences in product recommendations is unrelated to unbiased estimation of the preference model. Instead, it shows experimentally that the practical performance improvement of recommendation models using causal methodologies mainly results from the regularization effect of multi-task learning, which enriches the embedding space.

This research was supported by the Ministry of Science and ICT (National Research Foundation of Korea) through the Data Science Convergence Talent Development Project and the Basic Research Program (Mid-Career Research), "Research on Distributed Learning Using Language Models."

▶ Professor Jong-June Jeon, Department of Statistics; Wonhyung Shin, Ph.D. candidate

Ph.D. candidate Wonhyung Shin from the Department of Statistics and Data Science at the University of Seoul was the first author of the paper, and Professor Jong-June Jeon from the Department of Statistics was the corresponding author.

The research team stated, "We confirmed that robust generalization performance can be achieved through multi-task learning alone, without IPS reweighting, even when the dataset or platform changes," and added, "This research has immediate implications for large-scale recommendation engines, and the broader landscape of generative AI-based recommendation and search systems."

Meanwhile, the research team is conducting follow-up work to address popularity bias in sequential recommender systems from a causal inference perspective. They aim to improve generalization using real streaming logs and online A/B testing environments, while also developing lightweight, low-latency inference architectures for immediate deployment in large-scale production systems.