Investigation of Autoencoder Neural Network Accuracy for Computational Intelligence Methods to Estimate Missing Data

J. Mistry, F.V. Nelwamondo, and T. Marwala (South Africa)


Autoencoder Neural Network, Genetic Algorithm Optimi sation


Autoencoder neural networks are used in a computational intelligence method for estimating missing data. It is im portant to investigate what impact the accuracy of the au toencoder neural network has on the accuracy of the miss ing data estimation method. The data used for the inves tigation is a set of demographic properties of individuals obtained from the South African antenatal seroprevalence survey. Autoencoder neural networks were built using the Multi Layer Perceptron (MLP) Neural Network architec ture and together with a Genetic Algorithm Optimisation Method the missing data are estimated. Autoencoder neu ral networks with accuracies ranging from 80% to 100% estimate missing data with accuracies ranging from 55% to 65%. The autoencoder neural network must have a high ac curacy for the missing data to be correctly estimated. It is, however, important not to over-train the autoencoder neural network as this results in poor estimation of missing data.

Important Links:

Go Back