Hierarchical Active Learning with Label Proportions on Data Regions.

Area of Science:

• Machine Learning
• Data Mining
• Computational Biology

Background:

• Instance-based annotation for classification models is time-consuming and costly.
• Existing active learning methods often rely on extensive instance-level labeling.
• There is a need for efficient methods to reduce human annotation effort in model training.

Purpose of the Study:

• To propose a novel active learning framework that builds classification models from human-annotated regions.
• To address the challenge of limited initial regions by developing a hierarchical active learning (HAL) framework.
• To enhance the framework with a multi-hierarchy (forest) approach for more informative and diverse regions.

Main Methods:

• Developed a hierarchical active learning (HAL) framework that progressively divides the data space into sub-regions.
• Utilized learning from label proportions algorithms to train models using region labels and class proportions.
• Implemented a multi-hierarchy (forest) solution to build multiple shallower hierarchies of regions.
• Evaluated the framework on diverse classification datasets and a real-world user study in cancer survival analysis.

Main Results:

• Region-based active learning methods can effectively learn high-quality classifiers.
• The HAL framework significantly reduces the human annotation effort needed for building classification models.
• Demonstrated the framework's effectiveness on numerous classification datasets and a colorectal cancer survival analysis study.

Conclusions:

• The proposed region-based active learning framework offers a highly effective solution for reducing annotation costs.
• Active learning from regions provides a viable alternative to traditional instance-based annotation.
• The hierarchical and multi-hierarchy approaches enhance the efficiency and quality of model learning from limited labeled regions.