Active learning with drifting streaming data.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Labeling streaming data is costly and time-consuming.
• Active learning aims to minimize labeling effort for accurate models.
• Streaming data presents challenges due to concept drift and the need for continuous adaptation.

Purpose of the Study:

• To develop a theoretically supported framework for active learning from drifting data streams.
• To introduce novel active learning strategies that explicitly address concept drift.
• To improve model adaptability and accuracy in dynamic data environments.

Main Methods:

• Development of a theoretical framework for active learning in concept-drifting environments.
• Implementation of three distinct active learning strategies: uncertainty-based, dynamic allocation, and randomized search.
• Empirical evaluation of the proposed strategies on streaming datasets exhibiting concept drift.

Main Results:

• The proposed strategies effectively handle concept drift occurring anywhere in the instance space.
• Strategies demonstrate robust adaptation to unexpected changes in data distribution.
• Active learning methods show improved performance compared to conventional approaches in drifting data scenarios.

Conclusions:

• The presented framework and strategies offer a viable solution for active learning with concept drift.
• The developed methods enable more efficient and accurate learning from dynamic data streams.
• This research contributes to advancing active learning techniques for real-world streaming data applications.