Improving Classification Performance through an Advanced Ensemble Based Heterogeneous Extreme Learning Machines.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Extreme Learning Machine (ELM) is a rapid algorithm for single-hidden layer feedforward neural networks (SLFNs).
• ELM can overfit training data due to an excess of hidden nodes, potentially impacting generalization.
• Addressing overfitting and enhancing generalization are crucial for robust ELM performance.

Purpose of the Study:

• To introduce an Advanced ELM Ensemble (AELME) for improved classification performance.
• To mitigate overfitting issues inherent in standard ELM algorithms.
• To enhance prediction accuracy and generalization capabilities in classification tasks.

Main Methods:

• A heterogeneous ensemble approach combining Regularized-ELM, L2-norm-optimized ELM (ELML2), and Kernel-ELM.
• Ensemble construction involves training randomly selected ELM classifiers on data subsets obtained via random resampling.
• An objective function promoting diversity and accuracy within the ensemble guides the evolution process.

Main Results:

• AELME demonstrates superior prediction accuracy and generalization compared to Adaboost, Bagging, Dynamic ELM ensemble, data splitting ELM ensemble, and standard ELM ensemble.
• The proposed method requires a lower number of base classifiers.
• Validation on real-world benchmark datasets confirms the effectiveness of AELME.

Conclusions:

• The heterogeneous ensemble approach with AELME effectively addresses ELM overfitting and enhances generalization.
• AELME offers a robust solution for classification tasks, outperforming existing ensemble methods.
• The strategy of splitting data and incorporating diverse ELM classifiers leads to significant performance improvements.