Stacked Ensemble Deep Random Vector Functional Link Network With Residual Learning for Medium-Scale Time-Series...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Randomized neural networks, including deep random vector functional link (dRVFL) and ensemble dRVFL (edRVFL), show strong performance.
• Existing edRVFL architectures lack diversity and independent error correction capabilities within networks.

Purpose of the Study:

• To introduce novel ensemble deep random vector functional link network architectures with enhanced diversity and error correction.
• To develop a residual learning-based dRVFL (ResdRVFL) and an ensemble deep stacking network (SResdRVFL) that improve upon existing methods.

Main Methods:

• Combining stacked deep blocks and residual learning with the edRVFL framework.
• Proposing ResdRVFL, where deep layers correct shallow layer estimations, and incorporating a scaling parameter to control residual scaling and prevent overfitting.
• Developing SResdRVFL by aggregating multiple ResdRVFL blocks for ensemble learning.

Main Results:

• The proposed SResdRVFL model was evaluated on 28 diverse datasets.
• Comparative analysis showed SResdRVFL outperformed state-of-the-art methods in terms of average ranking and error rates.

Conclusions:

• The SResdRVFL architecture effectively addresses limitations in existing edRVFL models.
• The integration of residual learning and ensemble deep stacking provides superior performance and robustness across various machine learning tasks.