报告人：Peng Liuhua (彭柳华副教授，墨尔本大学)

报告题目：Imbalanced Learning with Beneficial Label Noise

报告摘要：Data imbalance and label noise are common factors hindering the classifier's performance in real-world datasets. Approaches to addressing imbalanced learning often involve resampling by adding or removing samples, which compromises data feature integrity and results in information loss or generative errors. Building upon theoretical studies of the impact of imbalance ratio on decision boundaries across various evaluation metrics, this paper proposes leveraging label noise to alter the biased decision boundaries via label flipping. By employing a carefully designed harmless asymmetric label noise model that re-balances the data by artificially flipping labels, it is uncovered that introducing appropriate label noise can enhance the performance of classifiers in imbalanced learning. In contrast to resampling methods, our novel approach re-balances the data while preserving the integrity of the feature space. We validated the superiority of this method over resampling approaches on simulated and real-world datasets with multiple classifiers. Overall, our work opens up a new avenue for addressing imbalanced learning, highlighting the potential advantages of balancing data through beneficial label noise.

报告时间：2024年7月2日（星期二）10:00-11:30

报告地点：东三十二楼115会议室

邀请人：潘灯

报告人简介：彭柳华博士，2017年获得爱荷华州立大学统计系统计博士学位。现受聘为墨尔本大学数学与统计学院高级讲师（等同于副教授）。彭博士的主要研究领域为高维数据分析，极值理论，大数据的分布式推断，bootstrap，以及非参数假设检验。彭博士在the Annals of Statistics，Journal of Royal Statistical Society SeriesB（JRSSB），Biometrika，Econometric Theory，Statistica Sinica等国际顶尖统计与计量期刊发表论文十余篇，多次应邀为顶尖统计期刊专家评审及在国际会议做邀请报告。