Active Learning by Querying Informative and Representative Examples.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Active learning (AL) aims to reduce data labeling costs by strategically selecting informative instances.
• Current AL methods often focus on either informativeness or representativeness, limiting performance.
• Combining these criteria in existing methods is often ad hoc.

Purpose of the Study:

• To develop a principled active learning approach that effectively combines informativeness and representativeness.
• To address the limitations of existing methods in selecting instances that are both informative and representative.
• To extend the approach to multi-label learning scenarios.

Main Methods:

• Developed QUIRE (Query Informativeness and Representativeness) based on a min-max view of active learning.
• Introduced a systematic method for measuring and combining informativeness and representativeness.
• Incorporated label correlations to extend the approach for multi-label learning.

Main Results:

• QUIRE systematically measures and combines informativeness and representativeness.
• The approach effectively handles both single-label and multi-label active learning.
• Experimental results demonstrate superior performance compared to state-of-the-art active learning methods.

Conclusions:

• The proposed QUIRE method offers a principled and effective solution for active learning.
• QUIRE enhances labeling efficiency by better selecting valuable data.
• The approach shows significant improvements in both single-label and multi-label learning tasks.