Online Active Learning Ensemble Framework for Drifted Data Streams

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Data stream classification presents challenges like high labeling costs and concept drift.
• Existing methods struggle with dynamic data changes and resource constraints.

Purpose of the Study:

• To develop an online active learning ensemble framework for drifting data streams.
• To address high labeling costs and concept drift effectively.

Main Methods:

• A hybrid labeling strategy combining an ensemble classifier (stable and dynamic classifiers) and active learning.
• Utilizing a multilevel sliding window model for dynamic classifier updates.
• Implementing an adaptive uncertainty sampling strategy with a dynamically adjusted decision threshold.

Main Results:

• Achieved precise prediction accuracy on synthetic and real datasets.
• Demonstrated effective handling of both gradual and sudden concept drift types.
• Showcased dynamic allocation of labeling costs based on concept drift.

Conclusions:

• The proposed framework efficiently manages labeling costs for drifting data streams.
• It provides a robust solution for accurate classification in dynamic environments.
• The method offers a cost-effective approach to active learning in data streams.