Detection of interaction articles and experimental methods in biomedical literature

Table 4 Comparison of feature set quality using t-test on stratified 10-fold CV subsets from the development set at best performing percentiles (per) of class 1.

f₁ @ per	CI	f₂ @ per	CI	p	EI
S @ 19	0.415, 0.480	P @ 35	0.397, 0.454	0.0379	0.022
B @ 13	0.460, 0.511	M @ 13	0.370, 0.461	0.0082	0.070
B @ 13	0.460, 0.511	S @ 19	0.415, 0.480	0.0273	0.038
W @ 20	0.491, 0.535	B @ 13	0.460, 0.511	0.0116	0.028
WS @ 22	0.510, 0.568	W @ 20	0.491, 0.535	0.0043	0.026
WMS @ 20	0.546, 0.577	WS @ 22	0.510, 0.568	0.0414	0.023
WMBS @ 17	0.549, 0.586	WS @ 22	0.510, 0.568	0.0179	0.029
WMPBS @ 17	0.558, 0.595	WS @ 22	0.510, 0.568	0.0122	0.038

Features considered are W (bag of words), M (MeSH), P (PPIscore), B (bigrams), S (syntactic) - for a detailed description see page 3. Interpret the rows as follows (e.g. row 4): Feature set W has a 95% confidence interval (CI) of [0.491, 0.535], feature set B has one of [0.460, 0.511]. According to a t-test for dependent samples, feature set W is significantly better than feature set B (df=9; p=0.0116). The expected improvement (EI) of the MCC measure is at least 0.028 (95% confidence level). Notice, that feature set PBMSW or BMSW are not significantly better than MSW. For the case of combinations of 2, 3 or 4 different feature sets, only the best performing ones were selected in this table.