.png){kind=icon}
.png){kind=icon}
Return
Resolving data imbalance in transfer learning: a simple random sampling approach
.png){kind=icon}
PRE
AI
DOI：10.1007/s11222-025-10811-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The prevalence of imbalanced data has been and remains a significant hurdle to the effective implementation of both Machine Learning and Transfer Learning techniques; although significant work has been devoted to resolving this problem in the context of traditional Machine Learning, the problem of imbalanced data in Transfer Learning has remained largely unaddressed. In this work, we fill this gap in the literature by investigating various random sampling techniques for resolving the data imbalance problem, with special focus on improving Neural Network and Transfer Learning performance in classification tasks on large imbalanced datasets. We first provide theoretical guarantees for the effectiveness of data balancing techniques in improving classification accuracy, as well as a theoretical guarantee of the performance superiority of Simple Random Undersampling compared to other undersampling techniques. We then verify these results using experimental performance of Neural Network classification and Transfer Learning fine-tuning on a simulated imbalanced dataset as well as five large real-world datasets in disease, bank marketing, income demographics, credit default, and image classification, observing that Simple Random Sampling is the empirically optimal technique for improving Machine Learning and Transfer Learning classification performance on imbalanced data. In all, our results provide strong evidence of Simple Random Sampling as the empirically optimal random data balancing technique for improving performance in Machine Learning and Transfer Learning.
Keywords:
Transfer Learning
Imbalanced Data Classification
Undersampling
Oversampling
Simple Random Sampling