Dynamic Ensemble Selection for Imbalanced Data Streams With Concept Drift.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Ensemble learning typically uses global performance for base classifier selection, which is inadequate for localized concept drift.
• Data streams often present class imbalance, negatively impacting ensemble learning accuracy for minority classes.

Purpose of the Study:

• To propose a dynamic ensemble selection for imbalanced data streams with concept drift (DES-ICD).
• To address the limitations of global performance criteria and class imbalance in ensemble learning for data streams.

Main Methods:

• A novel synthetic minority oversampling technique with adaptive nearest neighbors (AnnSMOTE) is developed to generate minority instances for new concepts.
• DES-ICD builds base classifiers on AnnSMOTE-balanced data chunks and merges them into a candidate pool.
• Optimal classifier combinations are selected based on neighborhood performance for query instances.

Main Results:

• The proposed DES-ICD method demonstrated superior performance over seven comparative methods.
• Experimental results on nine synthetic and five real-world datasets confirmed improved classification accuracy.
• The method accurately tracks new concepts in imbalanced data streams.

Conclusions:

• DES-ICD effectively addresses concept drift and class imbalance in data streams.
• The neighborhood-based selection strategy enhances the adaptability of ensemble models.
• The AnnSMOTE technique improves the handling of minority instances in evolving data.