Building multiclass classifiers for remote homology detection and fold recognition

BMC Bioinformatics

Table 8 Comparative results for the remote homology detection problem on dataset sf95.

	Ie et al [19]		Scaling Model		Best Model
	ZE	BE	ZE	BE	ZE	BE
Without Hierarchy Information
Ranking Perceptron	21.8	36.7	9.3	16.1	9.3	16.1
SVM-Struct	20.7	37.6	9.0	15.9	9.0	15.9
With Fold-level Nodes
SVM-Struct	20.4	37.5	11.2	19.6	10.1	19.3

The results for Ie et al were obtained from the supplementary website for the work [19], and represent the results obtained using the simple scaling model in their implementation. The results labeled "Scaling Model" correspond to the performance achieved by our two-level classifiers using the simple scaling model, whereas the results labeled "Best Model" correspond to the best performance achieved among the simple scaling, scaling & shift, and Cramer-Singer models. Both of these results were obtained from Table 2. All results were obtained by optimizing the balanced loss function. ZE and BE denote the zero-one error and balanced error percent rates respectively.

ISSN: 1471-2105