A Dictionary-Based Method for Classification with Universum Data
Zhiyong Che, Bo Liu, Yanshan Xiao · 2021
In fact, the collected examples included the third-class examples, they do not belong to positive samples or negative samples, which are referred as the Universum data. And Universum data can make better performance for the classifier. In this paper, a dictionary-based method for classification with Universum data is proposed to construct a unified model. In the proposed method, we embed the dictionary and Universum data to construct a unified framework, and the Universum data is introduced into the framework by the ɛ-insensitive loss. For the optimization, the SVD algorithm and gradient-based optimization methods are utilized to alternately optimize and update the dictionary, and the Lagrangian function is used to iteratively optimize the unified framework to obtain the classifier. Finally, extensive experiments are conducted on the benchmark datasets to evaluate the performance of the proposed U-DL method and baselines. The results have shown that the proposed U-DL method makes better performance than previous methods.