Concept Drift Adaptation by Exploiting Historical Knowledge.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Incremental learning with concept drift presents challenges for maintaining model accuracy over time.
• Existing ensemble methods often retrain historical models, raising questions about model preservation and utilization strategies.
• Effective handling of evolving data streams requires adaptive learning approaches.

Purpose of the Study:

• To propose a novel ensemble learning method, Diversity and Transfer-based Ensemble Learning (DTEL), for incremental learning with concept drift.
• To investigate optimal strategies for preserving and utilizing historical models in adaptive learning scenarios.
• To enhance the robustness of machine learning models against concept drift.

Main Methods:

• DTEL utilizes preserved historical models as initializations for new models.
• Transfer learning is employed to retrain historical models with newly arrived data.
• Emphasis is placed on preserving a diverse set of historical models, not solely those with high accuracy.

Main Results:

• Empirical studies on 15 synthetic and 5 real-world data streams demonstrated DTEL's effectiveness.
• DTEL outperformed four other state-of-the-art methods in handling concept drift.
• The method showed superior performance in adapting to changing data patterns.

Conclusions:

• DTEL offers a robust and effective approach to incremental learning in the presence of concept drift.
• Preserving model diversity alongside accuracy is crucial for adaptive ensemble methods.
• The proposed method provides a valuable contribution to the field of concept drift adaptation.