Continuous Graph Learning-Based Self-Adaptation for Multi-Stream Concept Drift.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Concept drift is a persistent challenge in nonstationary data streams, especially in multistream scenarios where interstream correlations change.
• Existing adaptation methods primarily focus on single streams, leaving a research gap in handling multistream concept drift.

Purpose of the Study:

• To propose a novel framework, Continuous Graph Learning-based self-adaptation framework (CGLM), for effectively addressing concept drift in multistream environments.
• To capture and adapt to changing interstream correlations dynamically.

Main Methods:

• Introduced a novel graph neural network (GNN) structure with a dynamic graph generator (AGG) to create adaptive correlation graphs from historical data.
• Implemented a self-adaptation process involving subgraph updating and continuous graph learning mechanisms for both non-drift and drift scenarios.
• Developed an adaptive diffusion graph attention module (ADGAT) to capture local correlation changes and adaptively update graph weights during concept drift.

Main Results:

• The proposed CGLM framework demonstrated superior performance compared to all baseline methods across three large-scale real-world datasets.
• CGLM maintained its effectiveness even when large-scale data was available for initial training.

Conclusions:

• CGLM offers a robust and effective solution for multistream concept drift adaptation by dynamically capturing and responding to interstream correlation changes.
• The framework's ability to self-adapt through continuous graph learning and adaptive attention mechanisms provides a significant advancement in handling complex nonstationary data environments.