计算机科学
数学
一致性(知识库)
算法
人工智能
数据挖掘
鉴定(生物学)
特征(语言学)
数据一致性
模式识别(心理学)
出处
期刊:2017 International Artificial Intelligence and Data Processing Symposium (IDAP)
[Institute of Electrical and Electronics Engineers]
日期:2026-09-01
卷期号:: 1-6
标识
DOI:10.1109/idap71869.2026.11727906
摘要
Missing values in tabular data are a fundamental issue that directly affects model performance in machine learning applications. Although the transformer-based TabPFN has recently demonstrated strong results on small-and medium-sized datasets, its behavior under missing data has not been systematically investigated. This study evaluates TabPFN’s tabular classification performance under controlled Missing Completely At Random (MCAR) missingness conditions. Missingness rates of 0%, 10%, 20%, and 40% were applied to seven datasets, including four real-world and three synthetic datasets. TabPFN was evaluated in its raw form (TabPFN_raw) and with median imputation (TabPFN_median) and missing-indicator (TabPFN_indicator) preprocessing strategies, and was compared with HistGradientBoosting, Random Forest, and Logistic Regression baselines. Based on the average relative decrease in Macro-F1 score at a 40% missingness rate, TabPFN_raw exhibited the lowest performance decline (8.81%), whereas the other models showed declines ranging from 11.88% to 15.23%. Friedman and Nemenyi analyses indicated statistically significant differences among the evaluated models and showed that TabPFN_raw significantly outperformed the HistGradientBoosting, Random Forest, and Logistic Regression baselines. These findings suggest that, under the evaluated MCAR conditions, TabPFN’s native missing-value handling is more robust than the examined simple preprocessing strategies.
科研通智能强力驱动
Strongly Powered by AbleSci AI