- The team developed a novel data imputation model, “U-VAE,” which accounts for missing value uncertainty

  The University of Seoul announced that a paper written by the research team led by Professor Jong-June Jeon from the Department of Statistics and Data Science has been accepted for the Main Technical Track of the Association for the Advancement of Artificial Intelligence (AAAI) 2026, the premier international conference on AI.

  AAAI is a globally authoritative academic conference on AI, registered as a BK Excellent International Academic Conference (recognized IF 4.0). This year’s conference saw high competitiveness, with approximately 31,000 papers submitted and only 4,167 accepted. The 2026 AAAI Conference is scheduled to be held at the Singapore Expo from January 20 to January 27, 2026.

  The accepted paper, “Impute Missing Entries with Uncertainty,” presents a novel, more sophisticated approach to address the problem of missing data, which inevitably occurs during actual data collection. Missing data have various causes, such as non-response of participants, omission of medical records, and measurement errors. They are not merely blank spaces, but fundamental sources of uncertainty that undermine the reliability of statistical analysis. While various single and multiple imputation techniques have been used, many methods have limitations: they either fail to adequately reflect this uncertainty or assume specific distributions, thereby missing the complex conditional distributions in real data.

▶ Framework of the Proposed Missing Value Imputation Methodology

  The U-VAE proposed by the research team is characterized by its design to non-parametrically approximate the entire conditional distribution of missing values. To achieve this, the Continuous Ranked Probability Score (CRPS) was introduced as a reconstruction loss function. Considering the practical constraint that actual values for missing data do not exist in real-world settings, a re-masking/un-masking technique was applied to learn diverse missing patterns. Furthermore, the team has proved an upper bound on the KL divergence, demonstrating how proximate the proposed imputation distribution is theoretically to the actual conditional distribution, thereby ensuring the model’s theoretical validity. Compared with existing models, the U-VAE showed superior performance in both single and multiple replacements across 11 real-world tabular datasets.

This study was supported by the Ministry of Science and ICT (National Research Foundation of Korea) through the “Research on Distribution Learning Using Language Models” under the Data Science Convergence Talent Development Project and the Basic Research Program (Mid-Career Research). The study included Ph.D. candidate Jaesung Lim from the Department of Statistics and Data Science at the University of Seoul and Professor Seunghwan An‬ from the Department of Information Telecommunication Engineering at Incheon National University as co-first authors, with Professor Jong-June Jeon serving as the corresponding author.

▶ Jong-June Jeon Professor , Jaesung Lim Ph.D. Candidate

  Professor Jong-June Jeon stated, “This research is significant because it goes beyond simply filling in values and proposes a method that models the uncertainty of missing values.”

  The research team is also conducting follow-up research to develop a methodology for imputing missing values using language models.