Concept Drift-Tolerant Transfer Learning in Dynamic Environments.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Traditional transfer learning methods struggle in dynamic environments prone to concept drift.
• Concept drift-tolerant transfer learning (CDTL) remains underexplored, particularly in adapting models to evolving target domains.

Purpose of the Study:

• To propose a hybrid ensemble approach for CDTL in nonstationary environments with streaming target data.
• To address the challenge of adapting target models and source domain knowledge to changing environments.

Main Methods:

• A class-wise weighted ensemble adapts target models to new data chunks by assigning independent weights per class.
• A domain-wise weighted ensemble integrates source and target models, selecting relevant knowledge.
• Adaptive weighted CORrelation ALignment (AW-CORAL) updates source models, minimizing domain discrepancy and reducing negative knowledge transfer.

Main Results:

• The proposed hybrid ensemble approach effectively handles concept drift in transfer learning.
• AW-CORAL successfully minimizes domain discrepancy while promoting positive knowledge transfer from source domains.
• Empirical studies on synthetic and real datasets validate the algorithm's effectiveness.

Conclusions:

• The developed hybrid ensemble method offers a robust solution for transfer learning in dynamic, nonstationary environments.
• The approach demonstrates significant improvements in adapting to concept drift by intelligently leveraging source domain knowledge.
• This work advances the field of CDTL by providing a practical and effective methodology.