Abstract:Over the past decades, technological advancements have included the production of a huge number of data streams. Data streams comprise large amounts of partially sequenced, infinite data. Changes in the statistical properties of the input data distribution, such as mean, variance and standard deviation or changes in their relationships with the target label, are referred to as concept drift. Concept drift and class imbalance poses challenges in maintaining accurate classification, adapting to evolving data patterns and effectively classifying minority classes. Addressing this problem requires techniques that handle both class imbalance and concept drift. If such problems are left unaddressed, they will hinder the learning model's performance. The proposed model in the paper uses adaptive synthetic sampling (ADASYN) to deal with class imbalances. The ADASYN method generates synthetic data by using a weighted distribution based on the severity of the minority class and hard to learn minority class samples. To adapt to concept drift, the performance-based diversified dynamic weighted ensemble is used subsequently. In addition, the cumulative sum statistical test is used to detect drift and, one of the ensemble's base learners, the self-organizing neural network model, which automatically creates a new layer when drift occurs and provide solution to catastrophic forgetting, performance-based pruning and ensemble evolution if more drift occurs. The proposed model's efficacy is assessed by utilizing a variety of state-of-the-art ensemble methods and seven datasets in a prequential test-then-train approach with single-pass learning. The results of the experiments show that the proposed model outperforms state-of-the-art ensemble methods.

Dynamic Ensemble of Ensembles in Nonstationary Environments

DE2: Dynamic Ensemble of Ensembles for Learning Nonstationary Data

Joint Sparse Learning for Classification Ensemble.

An Ensemble Model for 2D-Data Classification Based on Classical & Deep Learning Classifier

A novel classifier ensemble method with sparsity and diversity

Classification of High-Dimensional Evolving Data Streams Via a Resource-Efficient Online Ensemble

Classifier Ensemble Using A Heuristic Learning With Sparsity And Diversity

Subspace Ensembles for Classification

A New Ensemble Method for Multi-label Data Stream Classification in Non-stationary Environment

Dynamic Ensemble Selection for Imbalanced Data Streams with Concept Drift.

Ensemble Learning Method for Classification: Integrating Data Envelopment Analysis with Machine Learning

BatchEnsemble: An Alternative Approach to Efficient Ensemble and Lifelong Learning

CDA-PDDWE: Concept Drift-Aware Performance-Based Diversified Dynamic Weighted Ensemble for Non-stationary Environments

A distance-based weighting framework for boosting the performance of dynamic ensemble selection

Convex ensemble learning with sparsity and diversity

Dynamic ensemble selection methods for heterogeneous data mining

Adaptive Classifier Ensemble Method Based on Spatial Perception for High-Dimensional Data Classification

Sorting-Based Dynamic Classifier Ensemble Selection

Classification by evolutionary ensembles

Dynamic Ensemble Learning for Multi-label Classification

Study of Ensemble Algorithm and Its Application