Improving Generalization Capability of Extreme Learning Machine with Synthetic Instances Generation

In this paper, instead of modifying the framework of Extreme learning machine (ELM), we propose a learning algorithm to improve generalization ability of ELM with Synthetic Instances Generation (SIGELM). We focus on optimizing the output-layer weights via adding informative synthetic instances to the training dataset at each learning step. In order to get the required synthetic instances, a neighborhood is determined for each high-uncertainty training sample and then the synthetic instances which enhance the training performance of ELM are selected in the neighborhood. The experimental results based on 4 representative regression datasets of KEEL demonstrate that our proposed SIGELM obviously improves the generalization capability of ELM and effectively decreases the phenomenon of over-fitting.