Active Learning With Optimal Instance Subset Selection.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Traditional active learning (AL) relies on instance-based utility measures, evaluating instances independently.
• This independent evaluation may not yield optimal labeling subsets due to ignored instance interactions.
• Maximal individual instance utility does not guarantee an optimal subset for effective machine learning.

Purpose of the Study:

• To propose a novel active learning strategy, Active Learning with Optimal Subset Selection (ALOSS).
• To address the limitations of traditional AL by considering instance interactions and importance for subset selection.
• To enhance the efficiency and effectiveness of the active learning process through optimal subset identification.

Main Methods:

• ALOSS identifies optimal instance subsets by maximizing a utility value.
• It incorporates both individual instance importance and inter-instance disparity.
• This is achieved by constructing an instance-correlation matrix and formulating AL as a semidefinite programming problem.

Main Results:

• ALOSS effectively selects subsets that maximize overall utility.
• Experimental results show ALOSS outperforms existing state-of-the-art active learning approaches.
• The method demonstrates superior performance in selecting informative and diverse instance subsets.

Conclusions:

• ALOSS offers a more effective approach to active learning by considering instance relationships.
• The semidefinite programming formulation provides a robust framework for optimal subset selection.
• This work advances active learning by moving beyond simple instance-based utility to subset-level optimization.