A novel multi-task TSK fuzzy classifier and its enhanced version for labeling-risk-aware multi-task classification

Abstract

While the Takagi–Sugeno–Kang (TSK) fuzzy system has been extensively applied to regression, the aim of this paper is to unveil its potential for classification, of multiple tasks in particular. First, a novel TSK fuzzy classifier (TSK-FC) is presented for pattern classification by integrating the large margin criterion into the objective function. Where multiple tasks are concerned, it has been shown that learning the tasks simultaneously yields better performance than learning them independently. In this regard, the capabilities of TSK-FC are exploited for multi-task learning, through the use of a multi-task TSK fuzzy classifier called MT-TSK-FC. MT-TSK-FC is a mechanism that uses not only the independent sample information of each task, but also the inter-task correlation information to enhance classification performance. However, as the number of tasks increases, the learning process is prone to labeling risk, which can lead to considerable degradation in the performance of pattern classification. To reduce the risk, a labeling-risk-aware mechanism is proposed to enhance the performance of the MT-TSK-FC, thus leading to the development of the labeling-risk-aware multi-task TSK fuzzy classifier called LRA-MT-TSK-FC. Since the three proposed fuzzy classifiers – TSK-FC, MT-TSK-FC, and LRA-MT-TSK-FC – can all be implemented by solving the corresponding QP problems, global optimal solutions are guaranteed. Experiments on multi-task synthetic and real image datasets are conducted to comprehensively demonstrate the effectiveness of the classifiers.