Improved Feature-Selection Method Considering the Imbalance Problem in Text Categorization

Table 9

The accuracy comparison of ECE with nine feature-selection algorithms when the NB is used on 20-Newsgroups. The numbers in the parentheses are the difference of accuracy of the corresponding feature-selection algorithm from that of ECE.

Feature selections	400	800	1200	1600	2000
ECE	70.81 (—)	74.54 (—)	75.51 (—)	76.73 (—)	77.16 (—)
CHI	66.44 (−4.37)	72.36 (−2.18)	74.54 (−0.97)	75.62 (−1.11)	76.66 (−0.50)
DF	56.01 (−14.8)	63.02 (−11.52)	67.74 (−7.77)	70.64 (−6.09)	72.10 (−5.06)
GINI	72.81 (+2.00)	75.79 (+1.25)	76.78 (+1.27)	77.22 (+0.49)	77.76 (+0.60)
IG	46.47 (−24.34)	53.99 (−20.55)	59.42 (−16.09)	62.92 (−13.81)	65.55 (−11.61)
MI	20.77 (−50.04)	29.38 (−45.16)	46.90 (−28.61)	50.70 (−26.03)	56.21 (−20.95)
DIA	15.86 (−54.95)	24.76 (−49.78)	26.02 (−49.49)	27.14 (−49.59)	28.64 (−48.52)
CMFS	72.19 (+1.38)	74.76 (+0.22)	76.20 (+0.69)	77.72 (+0.99)	78.43 (+1.27)
OCFS	43.10 (−27.71)	51.05 (−23.49)	56.82 (−18.69)	61.41 (−15.32)	63.87 (−13.29)
DFPFS	65.33 (−5.48)	69.56 (−4.98)	71.48 (−4.03)	72.66 (−4.07)	73.39 (−3.77)