Co-Training Based Sequential Three-Way Decisions for Cost-Sensitive Classification

As a popular research topic in classification, face recognition has drawn great attention in recent years, which is researched mostly in verification and identification modes. In this paper, we focus on the identification problem. Traditionally, a low misclassification error rate remains a fundamental target of various face recognition systems. However, in some real applications, this target is far from reasonable on account of different misclassification costs and large numbers of unlabeled facial images. To solve this problem, we introduce a sequential three-way decisions model for cost-sensitive face recognition. Instead of achieving a low recognition error rate, we are concerned with seeking a minimum misclassification cost in each decision step. When labeled samples are insufficient, delayed decisions can be made, which make up the boundary region. To mitigate the problem of insufficient labeled samples, we use a co-training mechanism. With more labeled samples obtained, the delayed decisions will be converted to positive or negative decisions definitely. This sequential three-way decisions model is in accordance with the human decision-making process. Several experiments are performed to demonstrate the effectiveness of this research.