MOOD: Leveraging Out-of-Distribution Data to Enhance Imbalanced Semi-Supervised Learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Class imbalanced and partially labeled data are common in real-world scenarios, necessitating research in imbalanced semi-supervised learning (SSL).
• Naturally collected datasets often contain out-of-distribution (OOD) samples, which significantly degrade the performance of existing imbalanced SSL methods.
• Tail classes in imbalanced datasets are particularly vulnerable to performance drops when OOD data is present.

Purpose of the Study:

• To propose a novel imbalanced SSL method, mixup-OOD (MOOD), designed to leverage OOD data effectively.
• To enhance feature diversity for tail classes by utilizing OOD data as a valuable resource.
• To address the performance deterioration of SSL methods caused by the presence of OOD samples.

Main Methods:

• Filtering OOD data from unlabeled datasets.
• Fusing filtered OOD data with labeled data to augment feature representation, especially for tail classes.
• Developing a push-and-pull (PaP) loss function to differentiate between in-distribution (ID) and OOD samples, attracting ID instances and repelling OOD samples from class centroids.

Main Results:

• MOOD demonstrates superior performance compared to existing state-of-the-art imbalanced SSL methods.
• The proposed method shows robustness across datasets with varying degrees of class imbalance.
• MOOD maintains strong performance even with different proportions of OOD data present in the unlabeled set.

Conclusions:

• Leveraging OOD data, previously considered detrimental, can significantly benefit imbalanced SSL.
• The MOOD method offers a promising approach to improve the accuracy and robustness of machine learning models in complex, real-world data scenarios.
• The developed PaP loss is crucial for effectively separating ID and OOD data, enhancing model generalization.