Hierarchical Active Learning With Qualitative Feedback on Regions.

Area of Science:

• Machine Learning
• Data Science
• Human-Computer Interaction

Background:

• Training classification models typically requires extensive labeled data.
• Instance-based annotation is often inefficient and time-consuming for human annotators.

Purpose of the Study:

• To propose and evaluate a novel human supervision approach using data regions.
• To develop an efficient method for model learning with reduced annotation effort.

Main Methods:

• Introduced region-based labeling, where qualitative class proportions are assessed for data subspaces.
• Devised a hierarchical active learning process to recursively construct a region hierarchy.
• Integrated active learning strategies with human expertise for discriminative feature provision.

Main Results:

• Extensive experiments on nine datasets demonstrated the framework's effectiveness.
• A user study on colorectal cancer patient survival analysis validated the approach.
• The region-based active learning framework outperformed traditional instance-based methods.

Conclusions:

• Region-based active learning offers a more efficient and effective alternative to instance-based methods.
• This approach significantly reduces human annotation effort while maintaining model performance.
• The proposed framework shows promise for practical machine learning applications.