Classification methods play an important role in investigating crime in forensic research. Here we assess the relative performance of several classification methods, such as Logistic Regression （LR）, Linear Discrimi...Classification methods play an important role in investigating crime in forensic research. Here we assess the relative performance of several classification methods, such as Logistic Regression （LR）, Linear Discriminant Analysis （LDA）, Quadratic Discriminant Analysis （QDA）, Mixture Discriminant Analysis （MDA） and Classification Tree （CT） on glass identification data. We present a different approach to investigate the relative performance of the classifiers by invoking the tests of statistical significance and the receiver operating characteristic （ROC） curves in addition to estimating the probabilities of correct classification （PCC）. The area under the receiver operating characteristic curve （AUC）, the error rate and its 95% confidence interval are used to measure predictive power of these algorithms. Dimensionality reduction of data has been conducted using principal component analysis （PCA） and Fisher＇s linear discriminant analysis （FDA） and two major components were identified. Among all the classification methods mentioned above, the LDA and the QDA are observed to be statistically significant. The Box＇s M test （P〈0.0001）, which is used to test the homogeneity of covariance matrices, showed that the homogeneity of covariance could not be assumed for LDA. This suggests that for glass types, window and non-window, the QDA is superior to all methods. The CT, however, has been found to outperform FDA when all six categories of glass are considered.展开更多
文摘Classification methods play an important role in investigating crime in forensic research. Here we assess the relative performance of several classification methods, such as Logistic Regression （LR）, Linear Discriminant Analysis （LDA）, Quadratic Discriminant Analysis （QDA）, Mixture Discriminant Analysis （MDA） and Classification Tree （CT） on glass identification data. We present a different approach to investigate the relative performance of the classifiers by invoking the tests of statistical significance and the receiver operating characteristic （ROC） curves in addition to estimating the probabilities of correct classification （PCC）. The area under the receiver operating characteristic curve （AUC）, the error rate and its 95% confidence interval are used to measure predictive power of these algorithms. Dimensionality reduction of data has been conducted using principal component analysis （PCA） and Fisher＇s linear discriminant analysis （FDA） and two major components were identified. Among all the classification methods mentioned above, the LDA and the QDA are observed to be statistically significant. The Box＇s M test （P〈0.0001）, which is used to test the homogeneity of covariance matrices, showed that the homogeneity of covariance could not be assumed for LDA. This suggests that for glass types, window and non-window, the QDA is superior to all methods. The CT, however, has been found to outperform FDA when all six categories of glass are considered.
