Selecting Negative Samples for PPI Prediction Using Hierarchical Clustering Methodology

Table 3

Accuracy using the 6 most relevant features for three RBF-SVM models.

Datasets	Acc. Our proposal RBF-SVM	Acc. Rand. RBF-SVM	Acc. “balanced’’ RBF-SVM	% relative difference for our proposal versus “Rand’’ model	% relative difference for our proposal versus “balanced’’ model
Binary-GS	94,111	95,411	92,401	−1,381	1,817
Ito-core	72,059	81,195	52,571	−12,678	27,045
LC-multiple	93,750	95,924	93,517	−2,319	0,249
Uetz-screen	76,857	80,822	54,882	−5,159	28,592
Random negative dataset 1	72,211	38,353	6,537	46,888	90,947
Random negative dataset 2	71,951	37,937	37,444	47,274	47,959
𝑅^test_13	58,184	29,349	1,883	49,558	96,764
𝑅^test_23	63,596	30,150	1,882	52,591	97,041
𝑅^test_33	96,469	69,365	1,683	28,096	98,255
𝑅^test_43	62,221	31,061	1,522	50,080	97,554
𝑅^test_53	61,248	29,862	1,364	51,244	97,773
𝑅^test_63	64,992	33,120	1,702	49,040	97,381
𝑅^test_73	64,441	31,454	1,824	51,189	97,170
𝑅^test_83	94,705	67,821	1,702	28,387	98,203
𝑅^test_93	64,334	31,237	1,061	51,446	98,351

Acc. is the accuracy of the RBF SVM model. Our proposal RBF-SVM is the SVM model trained using the training set formed by the GSP set and the GSN set obtained using the proposed hierarchical clustering method. Rand. RBF-SVM is the SVM model trained using the training set where the GSN set was randomly selected. “balanced” RBF-SVM is the SVM model trained using the training set formed by the GSP set and the GSN set obtained using the approach to create a “balanced” negative set by Yu et al. [38]. % relative difference is the percentage of relative difference using “our proposal RBF-SVM” as basis.