FEATURE-SELECTION USING THE KALMAN FILTER FOR CLASSIFICATION OF MULTIVARIATE DATA

A Kalman filter is developed as a feature selection method and classif ier for multivariate data. Three near-infrared (NIR) data sets and a p ollution data set are analyzed. For the two most difficult data sets ( data sets 1 and 3), the Kalman filter successfully selects the wavelen gths which lead to very good results with a correct classification rat e (CCR) equal to one. These results are much better than the best resu lts obtained from regularized discriminant analysis (RDA) using Fourie r transform Fl, principal component regression (PCA) and univariate fe ature selection methods as the variable reduction methods. For the sec ond data set which consists of more than two classes, the Kalman filte r gives similar results (CCR=1) to those of RDA. For the pollution dat a set (data set 4), the Kalman filter gives similar results to partial least-squares (PLS) using fewer variables. The disadvantage of the Ka lman filter is that it needs more memory and more computing time than PLS. The potential hazards of overfitting have been considered.