A Multistream Concept Drift Handling Framework via Data Sharing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Concept drift, where data distribution changes over time, is a frequent problem in data stream mining.
• Insufficient data is a common challenge when dealing with concept drift.
• Existing methods often process multiple data streams independently, limiting their effectiveness.

Purpose of the Study:

• To propose a novel Multistream Concept Drift Handling Framework (MCDHF) to address insufficient data in concept drift.
• To leverage data sharing across multiple data streams for improved drift handling.
• To enhance prediction performance in data stream mining under concept drift.

Main Methods:

• Developed a Multistream Concept Drift Handling Framework (MCDHF) with fuzzy membership-based drift detection (FMDD) and adaptation (FMDA) components.
• Defined a stream fuzzy set with membership functions to quantify sample belongingness to a data stream.
• Implemented a data-sharing mechanism to transfer weighted data from non-drifting to drifting streams.

Main Results:

• The framework effectively detects concept drift occurrence and location across streams.
• Sharing weighted data from non-drifting streams successfully mitigates the insufficient data problem.
• Experimental results on synthetic and real-world data demonstrate significant improvements in prediction performance.

Conclusions:

• The proposed MCDHF effectively handles concept drift by utilizing inter-stream data sharing.
• The method addresses the critical issue of insufficient data in drifting streams.
• This approach offers a robust solution for improving the accuracy of data stream mining models.