Building multiclass classifiers for remote homology detection and fold recognition

BMC Bioinformatics

Table 5 Zero-one and Balanced error rates for the fold recognition problem optimized for the zero-one loss function.

		fd25		fd40
		ZE	BE	ZE	BE
MaxClassifier		42.0	60.3	44.4	64.6
Direct K-way Classifiers		42.8	59.4	43.0	62.7
Two-Level Approaches Without Hierarchy Information
	Scaling	39.9	52.9	32.2	50.6
Ranking Perceptron	Scale & Shift	38.4	51.3	27.3	44.8
	Crammer Singer	34.8	48.9	37.7	56.6
	Scaling	41.3	55.2	33.7	50.0
SVM-Struct	Scale & Shift	41.0	54.3	29.0	46.2
	Crammer Singer	36.6	49.4	32.5	49.6
Two-Level Approaches With Class-level Nodes
	Scaling	39.9	52.2	31.9	50.2
SVM-Struct	Scale & Shift	38.4	52.9	29.3	44.6
	Crammer Singer	39.2	51.8	32.8	52.9
Two-Level Approaches With Superfamily-level Nodes
	Scaling	39.5	53.9	31.3	48.8
SVM-Struct	Scale & Shift	39.9	53.4	31.3	48.4
	Crammer Singer	37.7	52.1	33.4	51.0
Two-Level Approaches With Superfamily-level and Class-level Nodes
	Scaling	39.2	52.2	27.3	41.0
SVM-Struct	Scale & Shift	39.9	53.9	28.4	44.1
	Crammer Singer	38.8	54.7	31.3	48.0

ZE and BE denote the zero-one error and balanced error percent rates respectively. The results were obtained by optimizing the zero-one loss function.

ISSN: 1471-2105