Robust Dual-Model Collaborative Broad Learning System for Classification Under Label Noise Environments
Wu Deng, Jiuru Shen, Jianming Ding, Huimin Zhao
Abstract
Broad learning system (BLS) has shown remarkable results in various machine learning tasks, but the fragility of BLS based on the minimum mean square error criterion under the label noise environment in Internet of Things (IoT) limits its application. To solve these problems, a dual-model collaborative label adaptive correction BLS based on the collaboration of KRPBLS and SKBLS, namely SC-KSBLS is proposed to significantly improve the robustness of BLS under label noise environment. First, a new BLS based on the Kernel risk-sensitive mean p-power (KRP) criterion, namely KRPBLS is developed to overcome the sensitivity by optimizing the output weight matrix and replacing the minimum mean square error criterion. Then, the label feature space is constructed, and the popular regularization strategy is designed to construct the SKBLS model with the collaborative training of KRP criterion and regularization to improve the learning ability of BLS. Finally, the interaction mechanism between KRPBLS and SKBLS is designed to effectively identify and correct label noise. The robustness theory of the SC-KSBLS has been analyzed and proven. The experimental results on nine datasets in IoT show that the SC-KSBLS significantly improves the robustness and performance, which provide an innovative solution under label noise environment.