N.U. Hlalele, F.V. Nelwamondo, and T. Marwala (South Africa)
Neuro-fuzzy, missing data, ANFIS, numerical methods
In this paper, the analysis of a neuro-fuzzy inference architecture for missing data imputation is presented. The background of the missing data problem is sketched along with the methods currently used to impute missing data in databases. Three datasets, namely, the HIV South African seroprevelance data, Puma 560 Robot Arm data and the letter image recognition data are used to investigate the ability of the inference system to estimate the missing data. The inference is found to have an accuracy of 60 % when imputing the age of the father in the HIV dataset and an almost zero error when imputing the angular acceleration of the robot arm data. The accuracy of imputing the width of the letter attribute box proved to be satisfactory with a correlation coefficient of 0.9636 indicating close correlation.
Important Links:
Go Back
