Abstract:  Extreme Learning Machine(ELM)is known to be a promising algorithm which makes learning speed extremely fast and thus attracts a lot of attentions from both academia and industry.However，due to the random determination of input－layer weights and biases for hidden nodes，it might generate some un－optimal parameters which may have a negative influence on the generalization performance and predicted robustness of ELM.To alleviate such weakness，a number of works have been proposed to further improve the generalization capability and stability of ELM.These existing methods principally focus on modifying or enhancing the structure of ELM and tend to suffer from over－fitting problem for paying excessive attention on matching the entire training dataset as well as increasing the model’s complexity substantially.In this paper，we proposed an effective learning algorithm(Extreme Learning Machine Generalization Improvement through Synthetic Instance Generation，SIGELM)to improve the generalization performance of ELM based on randomly generated synthetic instances.SIGELM does not need to modify the architecture of ELM model and uses the synthetic instances of which the distribution is approximately equal to the training instances with high uncertainty to optimize the output－layer weights of ELM.In order to obtain 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 the current updated ELM on the initial training dataset are selected in the neighborhood.The experimental results on four representative KEEL regression datasets demonstrated that SIGELM can significantly improve the generalization performance of ELM and meanwhile effectively control its over－fitting，which verified the feasibility and effectiveness of our proposed SIGELM algorithm.