LJCSNS International Joumal of Computer Science and Network Security, VOL 6 No5A, May 2006 4.2 Performance measures N-best accuracy was used for evaluation. We considered two kinds of criteria for accuracy,N=l"and N=l"meant that the most suitable category predicted by kNN and SVM corresponded to the original target document category, then a correct prediction was N=3 meant that at least one of three higher suitable categories that were predicted by kNN and SVM Binary Category corresponded to the original target document category, a correct prediction was considered Fig. 5. kNN: accuracy of"N=I' 4.3 Experimental Results Figure 5, Figure 6, Figure 7, and Figure 8 have the results obtained through the different feature selection methods we tested. The results using the knn classifier are presented in Figure 5 and Figure 6. The other results, Figure 7 and Figure 8 are used the svm classifier. The horizontal axis is the number of features, and the vertical axis is the accuracy score(\%)for N=I"andN=3 Additionally, we experimented the classification of kNN and SVM using an unbounded number of features The results of the experiments were that Accuracy scores of knn classification were N=l": $36.71%s andN=3" S61.2\%S. those of svm classification were N=1 $35.6\%S and"N=3":$61.0\%S. The accuracy of using ar unbounded number of features was lower than that of feature selections. For the result given above, feature Fig. 6. kNN: accuracy of N=3" selection was proven helpful in improving classification Furthermore, almost every result of accuracy scores was inary Category method Multivalued Category method In almost all cases, N=l"results of Figure 5 and Figure 7, revealed a higher accuracy for the Binary Category method than for the Multivalued Category method. Moreover, the Binary Category method atN=3", Figure 6 and Figure 8, was much more accurate than the Multivalued Category method with a fewer number of features. This helped to reduce the impact of noise and irrelevant data, and therefore 9 our feature selection method could reduce the computation a costs of classifying new documents without reducing accuracy For comparison of kNN(Figure 5, Figure 6)and SvM (Figure 7, Figure 8), their accuracy performance is approximate equivalent. This result was in agreement with [12] ber of feature Fig. 7. SVMIJCSNS International Journal of Computer Science and Network Security, VOL.6 No.5A, May 2006 21 4.2 Performance measures N-best accuracy was used for evaluation. We considered two kinds of criteria for accuracy, ``N=1'' and ``N=3''. • ``N=1'' meant that the most suitable category predicted by kNN and SVM corresponded to the original target document category, then a correct prediction was considered. • ``N=3'' meant that at least one of three higher suitable categories that were predicted by kNN and SVM corresponded to the original target document category, a correct prediction was considered. 4.3 Experimental Results Figure 5, Figure 6, Figure 7, and Figure 8 have the results obtained through the different feature selection methods we tested. The results using the kNN classifier are presented in Figure 5 and Figure 6. The other results, Figure 7 and Figure 8, are used the SVM classifier. The horizontal axis is the number of features, and the vertical axis is the accuracy score(\%) for ``N=1'' and ``N=3''. Additionally, we experimented the classification of kNN and SVM using an unbounded number of features. The results of the experiments were that Accuracy scores of kNN classification were ``N=1'' : $36.7\%$ and ``N=3'' : $61.2\%$, those of SVM classification were ``N=1'' : $35.6\%$ and ``N=3'' : $61.0\%$. The accuracy of using an unbounded number of features was lower than that of feature selections. For the result given above, feature selection was proven helpful in improving classification. Furthermore, almost every result of accuracy scores was Binary Category method > Multivalued Category method. In almost all cases, ``N=1'' results of Figure 5 and Figure 7, revealed a higher accuracy for the Binary Category method than for the Multivalued Category method. Moreover, the Binary Category method at ``N=3'', Figure 6 and Figure 8, was much more accurate than the Multivalued Category method with a fewer number of features. This helped to reduce the impact of noise and irrelevant data, and therefore our feature selection method could reduce the computation costs of classifying new documents without reducing accuracy. For comparison of kNN (Figure 5, Figure 6) and SVM (Figure 7, Figure 8), their accuracy performance is approximate equivalent. This result was in agreement with [12]. Fig. 5. kNN : accuracy of ``N=1'' Fig. 6. kNN : accuracy of ``N=3'' Fig. 7. SVM : accuracy of ``N=1