Smart adaptive ensemble model for multiclass imbalanced nonstationary data streams.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Real-time streaming data often presents simultaneous concept drift and class imbalance, degrading online machine learning model performance.
• Existing solutions primarily address these issues in binary class streams, with limited focus on multi-class scenarios.
• Ensemble learning, a common approach, struggles when new classifiers are not trained on relevant data reflecting concept changes.

Purpose of the Study:

• To propose a novel Smart Adaptive Ensemble Model (SAEM) for effectively handling concept drift and class imbalance in multi-class data streams.
• To enhance the robustness and accuracy of online machine learning models in dynamic environments.
• To address the limitations of current ensemble methods in adapting to evolving data concepts.

Main Methods:

• SAEM monitors feature-level data distribution changes to identify concept drift.
• A background ensemble is utilized to train new classifiers on data exhibiting changes.
• Dynamic class imbalance ratio weighting is applied to minority class instances to mitigate imbalance.

Main Results:

• The proposed SAEM demonstrated superior performance compared to state-of-the-art methods across eight diverse data streams.
• Significant average improvements were observed: 15.86% in accuracy, 20.35% in Kappa, 16.12% in F1-score, 15.58% in precision, and 16.42% in recall.
• Statistical analysis using the Friedman test confirmed significant performance differences across key metrics.

Conclusions:

• SAEM offers an effective and efficient solution for online learning applications facing concept drift and class imbalance in multi-class data.
• The model's adaptive nature and handling of imbalanced data contribute to its enhanced performance.
• The findings support SAEM's capability to maintain high model performance in dynamic, real-time data environments.