Stochastic configuration network ensembles with selective base models.

Area of Science:

• Machine Learning
• Computational Science

Background:

• Stochastic Configuration Networks (SCNs) offer rapid data modeling with theoretical learning power guarantees.
• Empirical SCN models show good practical performance, but deeper analysis of generalization is needed for effective ensemble creation.

Purpose of the Study:

• To develop a novel framework for building Stochastic Configuration Network (SCN) ensembles.
• To identify key factors influencing base model generalization performance in SCNs.
• To enhance the generalization capacities of SCN-based ensemble models.

Main Methods:

• Proposed a theoretical framework to analyze learner model generalization error under mild assumptions.
• Formulated a novel indicator using training errors, output weights, and hidden layer output matrix.
• Developed an algorithm to select optimal base models from a pool of randomized SCN learners.

Main Results:

• The proposed indicator effectively measures factors affecting generalization.
• The framework successfully identified subsets of appropriate base models for ensemble construction.
• Experimental results across various datasets demonstrated the advantages of the proposed ensemble method.

Conclusions:

• The novel framework enhances SCN ensemble construction by enabling informed selection of base models.
• The theoretical analysis and novel indicator provide a robust approach to improving generalization.
• The method shows significant advantages in building effective ensemble models for diverse applications.