An Adaptive Heterogeneous Online Learning Ensemble Classifier for Nonstationary Environments | JoVE Visualize

An Adaptive Heterogeneous Online Learning Ensemble Classifier for Nonstationary Environments

Tinofirei Museba¹, Fulufhelo Nelwamondo², Khmaies Ouahada²

¹Applied Information Systems Department, University of Johannesburg, Johannesburg, South Africa.

Computational Intelligence and Neuroscience|April 5, 2021

View abstract on PubMed

Summary

This study introduces Heterogeneous Dynamic Ensemble Selection based on Accuracy and Diversity (HDES-AD), a novel approach for machine learning in dynamic environments with concept drift. HDES-AD improves predictive performance by intelligently selecting diverse models, outperforming existing methods.

Area of Science:

Machine Learning
Data Science
Artificial Intelligence

Background:

Technological advances generate vast streaming data, often in nonstationary environments.
Concept drift, the evolution of data-generating processes, necessitates adaptive learning algorithms.
Existing methods focus on model updating, with less attention to selecting suitable model types for different drift scenarios.

Purpose of the Study:

To investigate the impact of heterogeneous online ensemble learning for predictive modeling in dynamic environments.
To propose a novel approach, Heterogeneous Dynamic Ensemble Selection based on Accuracy and Diversity (HDES-AD), for adaptive learning.
To enhance predictive performance in nonstationary environments by increasing diversity within ensembles.

Main Methods:

Developed HDES-AD, a heterogeneous ensemble approach using online dynamic ensemble selection.
HDES-AD interchanges different base models within an ensemble to boost predictive accuracy.
Employed diverse base learners to maintain ensemble diversity, addressing limitations of homogeneous ensembles.

Main Results:

HDES-AD demonstrated superior performance compared to homogeneous online ensemble methods (DDD, AFWE, OAUE).
The approach effectively enhances predictive performance in nonstationary environments.
Evaluated on both artificial and real-world datasets, confirming robustness and effectiveness.

Conclusions:

HDES-AD offers a significant advancement in handling concept drift in streaming data environments.
The proposed method provides a robust solution for adaptive learning in dynamic, nonstationary settings.
Heterogeneous ensembles with dynamic selection are crucial for maintaining performance as data distributions evolve.