Robust Extreme Learning Machine Based on an Improved p-Order Correntropy Loss Function

Extreme Learning Machine (ELM) has been widely applied in various pattern recognition and regression tasks due to its fast training speed and simple structure. However, conventional ELM exhibits poor generalization capability when dealing with noisy environments, lacking sufficient robustness and leading to performance degradation. To address this issue, this paper proposes an improvement to the objective function of ELM by incorporating the p-order correntropy as the loss function to enhance the model’s robustness against noisy data. Correntropy effectively measures the statistical similarity between predicted and true values, thereby suppressing the influence of outliers during the learning process in noisy environments and improving model stability. The experimental results show that across eight datasets and five different noise levels, the proposed method achieves the smallest error across all noise levels in five of the datasets. Moreover, as the noise level increases, the proposed method exhibits the smallest fluctuation range in error, demonstrating its stability and robustness.