Selective ensemble-based online adaptive deep neural networks for streaming data with concept drift.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Concept drift poses a significant challenge in streaming data mining, impacting real-time model convergence.
• Existing methods struggle with complex nonlinear problems and dynamic environments in streaming data classification.
• Maintaining model stability and adaptability is crucial for effective online learning.

Purpose of the Study:

• To propose a novel selective ensemble-based online adaptive deep neural network (SEOA) to address concept drift.
• To enhance the convergence and robustness of deep learning models in dynamic streaming data environments.
• To improve the classification accuracy for complex nonlinear streaming data problems.

Main Methods:

• Developed an adaptive depth unit combining shallow and deep features to control information flow.
• Implemented an ensemble of adaptive depth units as base classifiers, dynamically weighted by their loss.
• Introduced dynamic selection of base classifiers based on data fluctuations for a stability-adaptability balance.

Main Results:

• The proposed SEOA model demonstrates effective handling of various concept drift types.
• Experimental results show improved convergence for online deep learning models.
• The SEOA model exhibits good robustness and generalization capabilities in dynamic environments.

Conclusions:

• The SEOA model offers a robust solution for concept drift in streaming data mining.
• The adaptive depth units and ensemble strategy enhance model performance in dynamic settings.
• SEOA provides a promising approach for real-time classification of complex nonlinear streaming data.