Cost-sensitive sequential three-way decision modeling using a deep neural network

Three-way decision (3WD) models have been widely investigated in the fields of approximate reasoning and decision making. Recently, sequential 3WD models have attracted increasing interest, especially for image data analysis. It is essential to select an appropriate feature extraction and granulation method for sequential 3WD-based image data analysis. Among the existing feature extraction methods, deep neural networks (DNNs) have been considered widely due to their powerful capacity for representation. However, several important problems affect the application of DNN-based feature extraction methods to sequential 3WD. First, it takes a long time for a DNN to obtain an optimal feature representation. Second, most DNN algorithms are cost-blind methods and they assume that the costs of all misclassifications are the same, which is not the case in real-world scenarios. Third, DNN algorithms are two-way decision models and they cannot provide boundary decisions if sufficient information is not available. To address these problems, we propose a DNN-based sequential granular feature extraction method, which sequentially extracts a hierarchical granular structure from the input images. Based on the sequential multi-level granular features, a cost-sensitive sequential 3WD strategy is presented that considers the misclassification cost and test cost in different decision phases. Our experimental analysis validated the effectiveness of the proposed sequential DNN-based feature extraction method for 3WD.