Multi-task twin spheres support vector machine with maximum margin for imbalanced data classification

Multi-task learning (MTL) has been gradually developed to be a quite effective method recently. Different from the single-task learning (STL), MTL can improve overall classification performance by jointly training multiple related tasks. However, most existing MTL methods do not work well for the imbalanced data classification, which is more commonly encountered in our real life. The maximum margin of twin spheres support vector machine (MMTSVM) is proved to be an effective method for handling imbalanced data classification. Inspired by above study, this paper proposes a multi-task twin spheres support vector machine with maximum margin (MTMMTSVM) for imbalanced data classification. MTMMTSVM constructs two homocentric hyper-spheres for each task, meanwhile it explores the commonality to be shared and individuality of each task. Moreover, it introduces the maximum margin principle to separate the majority samples from the minority samples, thereby containing a linear programming problem (LPP) and a smaller quadratic programming problem (QPP). Compared with the latest multi-task algorithms, MTMMTSVM achieves superior g-mean and comparable accuracy on imbalanced datasets. Meanwhile, it dose not cost too much training time. Experiments have been conducted on five benchmark datasets, ten image datasets and one real Chinese wine dataset to explore the effectiveness of the MTMMTSVM. Finally, we employ a fast decomposition algorithm (DM) to handle the large-scale imbalanced problems more efficiently.

Automated summary

This paper proposes a multi-task twin spheres support vector machine with maximum margin (MTMMTSVM) for imbalanced data classification. It constructs two homocentric hyper-spheres for each task and explores the commonality and individuality of each task. Compared with the latest multi-task algorithms, MTMMTSVM achieves superior performance on imbalanced datasets and has a shorter training time.