The Use of Various Data Mining and Feature Selection Methods in the Analysis of a Population Survey Dataset

This paper reports the results of feature reduction in the analysis of a population based dataset for which there were no specific target variables. All attributes were assessed as potential targets in models derived from the full dataset and from subsets of it. The feature selection methods used were of the filter and wrapper types as well as clustering techniques. The predictive accuracy and the complexity of models based on the reduced datasets for each method were compared both amongst the methods and with those of the complete dataset. Analysis showed a marked similarity in the correlated features chosen by the supervised (filter) methods and moderate consistency in those chosen by the clustering methods (unsupervised). The breadth of distribution of the correlated features amongst the attribute groups was related in large part to the number of attributes selected by the given algorithm or elected by the user. Characteristics related to Health and Home, Paid and Volunteer Work and Demographics were the targets for which predictive accuracy was highest in both the reduced and full datasets. These attributes and a limited number of characteristics from the Learning, Social and Emotional attribute groups were important in clustering the population with Health and Home characteristics being most consistently important. Misclassification rates for models associated with most targets decreased with the use of subsets derived via filter methods but were increased for subsets derived using clustering methods.