Online Active Learning for Drifting Data Streams.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Streaming data classification faces challenges like concept drift, noise, and high labeling costs.
• Existing frameworks often fail to address these issues comprehensively.
• Efficiently labeling unlabeled data in dynamic environments remains a significant hurdle.

Purpose of the Study:

• To develop an active learning framework, CogDQS, for streaming data classification.
• To reduce the cost of manual annotation by intelligently selecting informative instances.
• To effectively manage concept drift and noise using principles of human memory cognition.

Main Methods:

• A dual-query strategy for instance selection based on local density and uncertainty.
• Leveraging Ebbinghaus's law of human memory (recall, fading period, memory strength) to manage concept drift and noise.
• Developing a policy to discern drift from noise and update outdated instances.

Main Results:

• CogDQS significantly reduces labeling costs compared to baseline methods.
• The framework demonstrates accurate and stable model performance across diverse data streams.
• Achieved good generalization ability even with gradual or abrupt drift and noise.

Conclusions:

• CogDQS offers an effective solution for cost-efficient and robust streaming data classification.
• The integration of memory cognition principles enhances the handling of dynamic data characteristics.
• The approach minimizes labeling, storage, and computation expenses while maintaining high accuracy.