A Robust Broad Learning System for Classification Problems in Label Noise Environments

In response to the problem of reduced performance of classification models due to label noise in real-world scenarios, this paper proposes a robust broad learning system that suppresses noise effectively and has parameter adaptive capabilities. The system is built within a broad learning framework characterized by fast learning ability and universal approximation ability. To enhance the model's ability to handle noise, a loss function is constructed using the kernel risk-sensitive mean p-power (KRP) criterion to train the output weights, which suppresses the influence of noisy samples in the loss function. The proposed model is theoretically proven to be robust. Experimental results on nine classification datasets demonstrate that the proposed model exhibits good generalization ability and robustness.