Sequential three-way decision based on multi-granular autoencoder features

Abstract

Autoencoder network is an efficient representation learning method. In general, a finer feature set obtained from autoencoder leads to a lower error rate and lower total misclassification cost. However, the network is usually trained for a long time to obtain a finer feature set, leading to a high time cost and total cost. To address this issue, a Sequential Three-Way Decision (S3WD) model is developed to balance the misclassification cost and the time cost in autoencoder based classifications and decisions. To implement the tradeoff strategy, it is necessary to extract a multi-granular feature set. In the network, the associated discriminative information in the extracted features increases with training epochs, which constructs a multi-granular feature structure. An autoencoder-based multi-granular feature description definition is presented. Based on the definition, an autoencoder composed of restricted Boltzmann machines is adopted to extract the multi-granular features. Then, a new cost-sensitive S3WD model is proposed, which aims to find the optimal granule level with the lowest total cost. Finally, the experiments demonstrate the effectiveness of the proposed approaches.