ISBN: 978-981-18-3959-7 DOI: 10.18178/wcse.2022.06.055
Feature Completion Using Correlation-Preserved Autoencoder
Abstract—Missing value existing in various datasets always causes tremendous obstacles to data mining in the real world. In recent years, autoencoder has been a popular deep learning model in data imputation due to the simple structure and efficient training period. In this paper, we developed a model that combined advantages of multiple imputations using denoising autoencoder (MIDA) and tracking-remove autoencoder (TRAE), integrating the idea of tracking-remove into MIDA. We introduce KNN to pre-impute the dataset after the input was denoising, and then we put the pre-imputed input data into both MIDA and MIDA with tracking-remove and ensemble the outputs by linear combination corresponding to the “multiple imputation” thought. The model called correlation-preserved autoencoder (CPAE) is applied to the completion of brain tissue feature data in ADNIMERGE (ADNI database). Experiments show that CAPE has a better performance than MIDA, TARE, and other autoencoders.
Index Terms—data imputation; Alzheimer's disease; autoencoder; KNN; data fusion
Li Sun, Tao Liu, Jiyun Li
School of Computer Science and Technology, Donghua University Shanghai, CHINA
Cite:Li Sun, Tao Liu, Jiyun Li, "Feature Completion Using Correlation-Preserved Autoencoder, " Proceedings of 2022 the 12th International Workshop on Computer Science and Engineering (WCSE 2022), pp. 389-394, June 24-27, 2022.