Exploiting Meta-Learned Confidences for Imbalanced Multilabel Learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Multilabel learning assigns multiple labels per sample, often overlooking label correlations.
• Class imbalance is a significant challenge in multilabel datasets, leading to biased classifiers.
• Existing research on imbalanced multilabel learning (IMLL) is limited.

Purpose of the Study:

• To propose a novel ensemble learning method for effective imbalanced multilabel learning.
• To leverage meta-learned confidences for improved label correlation and bias reduction.
• To enhance the performance of multilabel classifiers on imbalanced datasets.

Main Methods:

• Developed a meta-confidence ensemble (MCE) method for IMLL.
• Utilized out-of-bag samples to iteratively generate and fuse meta-confidences with original features.
• Employed a sparse projection method to prevent overfitting and enhance ensemble diversity.
• Implemented a calibrated plurality vote for final predictions on unseen samples.

Main Results:

• Meta-confidences effectively capture high-order label correlations.
• Meta-confidences aid in calibrating predictions, mitigating class imbalance issues.
• The proposed MCE method demonstrated superior performance in extensive experiments on IMLL tasks.

Conclusions:

• MCE offers a robust solution for imbalanced multilabel learning by exploiting meta-learned confidences.
• The method successfully addresses both label correlation and class imbalance challenges.
• MCE provides a significant advancement in the field of multilabel classification.