Protein-protein interaction extraction with feature selection by evaluating contribution levels of groups consisting of related features

Table 9 Comparison of our three approaches (BEST1G, U3G, O2G) with other systems

Corpus		LLL			HPRD50			IEPA			AIMed
(%)		P	R	F	P	R	F	P	R	F	P	R	F
	BEST1G	74.7	82.2	76.5	72.5	72.6	71.6	67.7	71.3	69.2	50.8	40.9	45.1
	U3G	74.6	80.7	75.9	72.5	72.6	71.6	68.3	71.7	69.8	49.5	40.5	44.4
Feature-	O2G	74.7	82.2	76.5	73.0	74.3	72.6	68.1	71.3	69.5	50.2	39.8	44.3
based	Landeghem et al. [2]	72.0	73.0	73.0	60.0	51.0	55.0	64.0	70.0	67.0	49.0	44.0	46.0
methods	Liu et al. [1]			78.1			64.9			62.1	63.4	48.8	54.7
	Yakushiji et al. [19]										33.7	33.1	33.4
Kernel-	Airola et al. [3]	72.5	87.2	76.8	64.3	65.8	63.4	69.6	82.7	75.1	52.9	61.8	56.4
based	Miwa et al. [4]	77.6	86.0	80.1	68.5	76.1	70.9	67.5	78.6	71.7	55.0	68.8	60.8
methods	Tikk et al. [11]	69.3	93.2	78.1	62.2	87.1	71.0	58.8	89.7	70.5	50.1	41.4	44.6
	Qian et al. [5]			84.6			68.8			69.8	59.1	57.6	58.1
Co-occurrence	Airola et al. [3]	55.9	100.0	70.3	38.9	100.0	55.4	40.8	100.0	57.6	17.8	100.0	30.1
Rule-based	RelEx [13, 17]	82.0	72.0	77.0	76.0	64.0	69.0	74.0	61.0	67.0	40.0	50.0	44.0
methods	Kabiljo et al. [18]	76.7	40.2	52.8	52.0	55.8	53.8	66.2	51.3	57.8	29.1	52.9	37.5

Performance comparison of our three approaches (BEST1G, U3G, O2G) with other related research on four corpora: LLL, HPRD50, IEPA, and AIMed. Co-occurrence and rule-based methods results are also listed as a baseline. Precision (P), recall (R), and F-score (F) values are shown by percentage (%). Bold typeface shows best results of feature-based and kernel-based methods per corpus in terms of precision, recall, and F-score

ISSN: 1471-2105