A rough margin-based multi-task v-twin support vector machine for pattern classification

Twin support vector machine (TSVM) has attracted significant attention in recent years, but it is suitable for solving the single-task learning (STL) problems. It trains each task independently and neglects the relationships among all tasks. Conversely, multi-task learning (MTL) explores the shared information between multiple correlated tasks, which obtains a better classifier than STL. Nevertheless, the existing multi-task twin support vector machines give the same penalties to the misclassified samples. In fact, the misclassified samples play a different role in generating separating hyperplane. Motivated by above studies, we put forward a rough margin-based multi-task v-twin support vector machine (rough MT-v-TSVM) in this paper. The proposed rough MT-v-TSVM gives different penalties to the misclassified samples depending on their positions. It not only takes full advantage of rough v-TSVM, but also discovers the commonality among tasks and maintains the individuality of each task. Therefore, compared with the state-of-the-art algorithms, our method yields better classification performance. In addition, we apply it to Chinese wine dataset to verify the effectiveness. Finally, the related extensions are further discussed, especially a fast SMO-type decomposition method (SDM) is introduced to handle relatively large-scale problems for acceleration. Comprehensive experiments are conducted on eleven benchmark datasets and an image dataset. The results demonstrate that our proposed algorithm can avoid over-fitting and achieve better classification accuracy, meanwhile it does not increase computational time compared with DMTSVM and MT-v-TSVM. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

TSVM is suitable for STL problems, while MTL explores shared information between multiple tasks for better classification. The proposed rough MT-v-TSVM assigns different penalties to misclassified samples based on their positions, combining the advantages of rough v-TSVM and preserving the individuality of tasks.