Hierarchical Active Learning with Proportion Feedback on Regions.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Supervised classification models typically require extensive human annotation, which is time-consuming and expensive.
• Reducing annotation costs is crucial for practical machine learning model development.
• Existing methods often rely on instance-level labeling, posing scalability challenges.

Purpose of the Study:

• To develop a cost-effective annotation strategy for classification models.
• To introduce region-based annotation as an alternative to instance-based annotation.
• To propose an active learning framework that leverages region-based feedback for efficient model training.

Main Methods:

• Exploration of region-based annotation, where a hyper-cubic subspace of the feature space is labeled with a proportion of positive class instances.
• Development of an active learning framework that hierarchically divides the input space.
• Iterative splitting of regions with the highest potential for classification model improvement.

Main Results:

• The proposed region-based active learning approach demonstrates effectiveness in learning classification models.
• Experiments show that this method can outperform existing active learning strategies.
• The approach is particularly advantageous in scenarios with limited labeling budgets.

Conclusions:

• Region-based annotation offers a promising new direction for active learning in classification.
• This method significantly reduces annotation effort and cost while maintaining or improving model performance.
• The hierarchical division and iterative refinement strategy enables learning accurate classifiers from proportional region feedback.