Application of Imbalanced Data Classification Quality Metrics as Weighting Methods of the Ensemble Data Stream...

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Massive data generation poses challenges for processing and classification.
• Concept drift and class imbalance are significant issues in data stream analysis.
• Existing methods like AWE and AUE struggle with imbalanced binary classification.

Purpose of the Study:

• To modify AWE and AUE for enhanced binary classification of imbalanced data streams.
• To improve adaptation to time-varying data distributions (concept drift).
• To address the challenge of disproportionate class representation in data.

Main Methods:

• Incorporated aggregate metrics (F1-score, G-mean, balanced accuracy) into classifier weight calculations.
• Investigated the impact of data sampling techniques on algorithm performance.
• Modified existing AWE and AUE stream processing solutions.

Main Results:

• Proposed modifications improved classification quality on imbalanced data.
• Enhanced algorithms demonstrated superior performance compared to base AWE/AUE.
• Experimental evaluation confirmed effectiveness against other imbalanced data stream solutions.

Conclusions:

• Modified AWE and AUE algorithms offer improved classification accuracy for imbalanced data streams.
• The inclusion of aggregate metrics effectively enhances classifier weighting and prediction.
• The study provides effective solutions for real-world problems like fraud and network attack detection.