Takagi-Sugeno-Kang Fuzzy System Towards Label-scarce Incomplete Multi-View Data Classification

Abstract

Fuzzy systems have shown a powerful ability in multi-view learning. However, most existing multi-view fuzzy systems require a large amount of labeled and complete multi-view data, which is a demanding requirement for real-world applications. To address these challenges, a novel transductive semi-supervised incomplete multi-view TSK fuzzy system (SSIMV_TSK) is proposed. First, to deal with the scarcity of labeled data, we introduce transductive learning, in which pseudo labels is assigned for unlabeled data. Then, to handle the incomplete multi-view data simultaneously, we integrate missing view imputation, pseudo label learning, and fuzzy system training as a single process to enable mutual learning. Furthermore, two new mechanisms – bidirectional structural preservation of instances and labels, and the adaptive multiple alignment collaborative learning – are proposed to further improve the robustness. Compared with the existing methods, the highlights of the proposed SSIMV_TSK are the ability to handle label scarcity and incomplete multi-view data simultaneously, the integration of missing view imputation and modeling as one process to make the imputed views more pertinent for classification; and the introduction of two new mechanisms to improve the quality of the learned pseudo labels and the imputed views. Experimental studies show that SSIMV_TSK significantly outperforms the state-of-the-art methods.