Hierarchical Active Learning with Group Proportion Feedback.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Supervised learning models require substantial human annotation, which is time-consuming and expensive.
• Reducing annotation costs is crucial for practical implementation of classification models.
• Existing methods often struggle to balance cost and model performance effectively.

Purpose of the Study:

• To develop a novel approach for training classification models with reduced human annotation effort.
• To introduce a hierarchical active learning framework that leverages group-based feedback.
• To decrease the overall cost associated with labeling data for machine learning.

Main Methods:

• Proposed a new active learning strategy that learns from groups (subpopulations) of data instances.
• Introduced human feedback on groups, providing estimated class proportions (a number in [0,1]).
• Developed a hierarchical framework to progressively refine models using annotated subpopulations.

Main Results:

• The proposed group-based active learning method demonstrated competitive performance.
• The approach significantly outperformed existing methods in reducing human annotation costs.
• Extensive experiments across multiple datasets validated the effectiveness of the framework.

Conclusions:

• The hierarchical active learning framework effectively reduces annotation costs in classification model training.
• Learning from group proportions offers a viable and efficient alternative to instance-level labeling.
• This method provides a practical solution for cost-sensitive machine learning applications.