Deep learning currently provides the best solutions in various industries involving tremendous data, such as object recognition and intrusion detection. In deep learning models, the quality and volume of data are two of the factors that determine task performance. This study concentrates on utilizing high-quality data to simultaneously improve the efficiency and accuracy of deep networks. This paper proposes a two-stage learning model that aims to generate high-quality data with reduced features during the first stage. Then, the selected data subset is regarded as the input in the second stage, i.e., the deep learning stage. However, most existing feature selection methods neglect the combination effect induced by inte-grated feature subsets. A correlation information entropy-based approach is developed to evaluate the integrated non-linear subspace. Experiments are carried out on six well-known classification datasets. The results indicate that our proposed two-stage learning model performs better than the compared high-dimensional deep learning models in speeding up the learning process and improving classification accuracy. Moreover, our developed feature selection method outperforms state-of-the-art feature selec-tion techniques in terms of time consumption and classification accuracy when combined with three deep learning models.(c) 2022 Elsevier B.V. All rights reserved.
展开▼
