Universum Least Square Twin Extreme Learning Machine
Qing Feng Wu, Jie Yang, Keyun Han, Yongfei Ye · 2025
Universum data, consisting of samples not belonging to any class in the learning task, can serve as prior knowledge for training models. Based on the advantages of Universum data, we propose a Universum least square twin extreme learning machine (ULSTELM). The method incorporates prior knowledge from unlabeled Universum data into the model framework to enhance the generalization ability. By leveraging prior knowledge, the classification accuracy can be significantly improved. ULSTELM is also designed to address problems with nonlinear scenarios. To validate the effectiveness of ULSTELM, we perform a series of experiments on datasets encompassing both Universum data and solely labeled data, including synthetic datasets, benchmark datasets, and MNIST datasets. The experimental results demonstrate that incorporating Universum data significantly improves generalization ability and classification accuracy, especially when fewer labeled samples exist.