Comparing Pattern Recognition Feature Sets for Sorting Triples in the FIRST Database

Pattern recognition techniques have been used with increasing success for coping with the tremendous amounts of data being generated by automated surveys. Usually this process involves construction of training sets, the typical examples of data with known classifications. Given a feature set, along with the training set, statistical methods can be employed to generate a classifier. The classifier is then applied to process the remaining data. Feature set selection, however, is still an issue. This paper presents techniques developed for accommodating data for which a substantive portion of the training set cannot be classified unambiguously, a typical case for low-resolution data. Significance tests on the sort-ordered, sample-size-normalized vote distribution of an ensemble of decision trees is introduced as a method of evaluating relative quality of feature sets. The technique is applied to comparing feature sets for sorting a particular radio galaxy morphology, bent-doubles, from the Faint Images of the Radio Sky at Twenty Centimeters (FIRST) database. Also examined are alternative functional forms for feature sets. Associated standard deviations provide the means to evaluate the effect of the number of folds, the number of classifiers per fold, and the sample size on the resulting classifications. The technique also may be applied to situations for which, although accurate classifications are available, the feature set is clearly inadequate, but is desired nonetheless to make the best of available information.