Partial correlation based variable selection approach for multivariate data classification methods

Selection of meaningful features characterizing the given set of system observations into distinct classes is crucial in all classification problems. A new significant attribute selection method based on partial correlation coefficient matrix (PCCM) is proposed. Many well studied representative classification data sets with different sizes and types are selected for investigating the performance. Linear Discriminant Analysis (LDA) combined with dimensional reduction techniques is employed as benchmark classifier to validate the new approach. The correlated attributes are arranged in order of their significance to multi-group data classification performance before applying the classification algorithm. Varying number of attributes are retained for the final analysis after PCCM based selection and progressive prediction accuracies are used to compare existing algorithms with the proposed feature selection algorithm. LDA results after PCCM based attribute selection show improvement in prediction efficiencies. It is shown that the PCCM based method is a better variable selection method compared to existing methods for obtaining the optimum set of predictor variables.