Hierarchical Active Learning with Overlapping Regions.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Instance annotation for classification models is time-consuming and costly.
• Reducing annotation effort is critical for developing effective machine learning models.
• Existing methods often require extensive human input for labeling individual data points.

Purpose of the Study:

• To explore region-based feedback as a novel human feedback mechanism.
• To develop methods for reducing annotation costs in machine learning.
• To investigate the effectiveness of learning from label proportions using region-based feedback.

Main Methods:

• Defined regions as hypercubic subspaces representing data subpopulations.
• Utilized learning from label proportions (LLP) algorithms with region-based feedback.
• Proposed a hierarchical active learning (HAL) approach to build concise region hierarchies.
• Developed a parallel hierarchy growth strategy to identify informative hierarchies.

Main Results:

• Region-based feedback significantly reduces human annotation effort.
• Effective classifiers can be learned with very few and simple region queries.
• Hierarchical active learning efficiently identifies relevant data subpopulations.
• Parallel hierarchy expansion improves the informativeness of the learning process.

Conclusions:

• Region-based feedback offers a highly effective alternative to instance annotation.
• The proposed hierarchical active learning method minimizes query complexity and number.
• This approach substantially lowers the cost and time associated with building classification models.