Iterative nearest neighborhood oversampling in semisupervised learning from imbalanced data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Transductive graph-based semisupervised learning relies on propagating labels through graphs.
• Existing methods are sensitive to imbalanced labeled datasets, skewing class boundaries.
• Majority classes can disproportionately influence classification outcomes in imbalanced scenarios.

Purpose of the Study:

• To propose a simple and effective approach to mitigate the impact of data imbalance in semisupervised learning.
• To enhance the performance of graph-based semisupervised learning methods on imbalanced datasets.
• To improve the accuracy of minority class predictions.

Main Methods:

• Iteratively selecting unlabeled samples to augment minority classes.
• Creating a balanced labeled dataset for subsequent learning.
• Utilizing graph-based semisupervised learning techniques on the balanced dataset.

Main Results:

• The proposed approach effectively alleviates the negative effects of data imbalance.
• Experiments on UCI and MNIST datasets demonstrate superior performance compared to state-of-the-art methods.
• Improved classification accuracy, particularly for minority classes.

Conclusions:

• The proposed iterative data balancing method is a simple yet powerful technique for imbalanced semisupervised learning.
• This approach offers a significant improvement over existing methods for handling imbalanced datasets.
• The findings have implications for real-world applications requiring robust classification with skewed data distributions.