Values Deletion to Improve Deep Imputation Processes
Most machine learning algorithms are based on the assumption that available data are completely known, nevertheless, real world data sets are often incomplete. For this reason, the ability of handling missing values has become a fundamental requirement for statistical pattern recognition. In this article, a new proposal to impute missing values with deep networks is analyzed. Besides the real missing values, the method introduces a percentage of artificial missing (‘deleted values’) using the true values as targets. Empirical results over several UCI repository datasets show that this method is able to improve the final imputed values obtained by other procedures used as pre-imputation.
KeywordsImputation Method Deep Neural Network Imputation Procedure Imputation Technique Noisy Version
- 1.García-Laencina, P.J., Sancho-Gómez, J.L., Figueiras-Vidal, A.: Pattern classification with missing data: a review. Neural Comput. Appl. 9(1), 1–12 (2009)Google Scholar
- 5.Bengio, Y.: Learning deep architectures for AI. Technical report, Dept. IRO, Universite de Montreal (2009)Google Scholar
- 7.Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Adv. Neural Inf. Process. Syst. 19, 153 (2007). (NIPS06), (B. Scholkopf, J. Platt, and T. Hoffman, eds.)Google Scholar
- 8.Lichman, M.: UCI machine learning repository (2013). http://archive.ics.uci.edu/ml