Exploring Representativeness and Informativeness for Active Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Active learning iteratively refines training datasets to enhance classifier performance across various applications.
• Existing active sampling methods often rely on specific data structures and struggle to fuse representativeness and informativeness criteria universally.
• There is a need for a general active learning framework that can combine these criteria without data structure assumptions.

Purpose of the Study:

• To propose a general active learning framework that effectively fuses sample representativeness and informativeness.
• To develop a method for selecting query samples that are both representative of unlabeled data and diverse within labeled data.
• To introduce a practical algorithm within this framework that demonstrates superior performance.

Main Methods:

• A novel framework inspired by two-sample discrepancy problems, using triple measures to ensure sample representativeness and diversity.
• Leveraging an uncertainty measure for the informativeness criterion, adaptable to different implementations.
• Proposing a specific algorithm using radial basis functions, estimated probabilities, and a modified best-versus-second-best strategy.

Main Results:

• The proposed framework successfully fuses representativeness and informativeness without prior data assumptions.
• The practical algorithm consistently outperformed state-of-the-art active learning methods on benchmark datasets.
• Experimental results validate the effectiveness of the developed triple measures and uncertainty measure.

Conclusions:

• The proposed general active learning framework offers a robust approach to sample selection.
• The practical algorithm derived from this framework significantly improves classification performance.
• This work provides a versatile solution for active learning challenges across diverse datasets.