Comparison of Support Vector Machines to Other Classifiers Using Gene Expression Data

ABSTRACT

Support vector machines (SVMs) was shown to outperform Fisher's linear discriminant analysis and two classification trees (C4.5 and MOC1) in binary classification of microarray gene expression data (MGED) (Brown et al., 2000; Furey et al. 2000). However, multiclass classification is more commonly encountered in identifying tumor subtypes using MGED. Using MGED, Dudoit et al. (2002) showed that diagonal linear discriminant analysis (DLDA) outperformed other linear and quadratic discriminants, nearest neighbor, and classification trees with univariate splits. It is of interest, therefore, to compare performance of SVMs to DLDA and the latest two classification trees with linear splits, which performered better than trees with univariate splits, in multiclass classification of MGED.

Furthermore, the performance of SVMs with different types of kernels were studied by three types of multiclass MGED. Finally, we investigate how irrelevant and correlated variables (features) influence the performance of the three classifiers. Some suggestions are made for multiclass classification of MGED.

Keywords:

• Classification
• Machine learning
• Microarray gene expression data
• Support vector machines
• Tumor class

Mathematics Subject Classification:

• 62P10
• 62H30

Acknowledgments

We thank Chi Lee for writing codes for Tables 2 3 4 5 and Drs. Y. C. Chang, C. J. Lin, and Sandrine Dudoit for helpful discussions. Thanks also go to the anonymous referee for constructive suggestions. Grace S. Shieh's research was supported in part by NSC grants 90-2118-M001-022 and 91-3112-P001-037.

Notes

d: the tuned value of the degree parameter in polynomial kernel.

d: the tuned value of the degree parameter in polynomial kernel.

d: the tuned value of the degree parameter in polynomial kernel.

L: SVMs with the linear kernal.

R: SVMs with the radial kernal.

a(0): only the top-10 ranked genes used.

b(0): only the top-10 ranked genes used.

L: QUEST with the linear split.

U: QUEST with the univariate split.

a(0): only the top-10 ranked genes used.

b(0): only the top-10 ranked genes used.

L: CRUISE with the linear split.

1D: CRUISE with the univariate 1D split.

2D: CRUISE with the univariate 2D split.

a(0): only the top-10 ranked genes used.

b(0): only the top-10 ranked genes used.

a(0): only the top-10 ranked genes used.

b(0): only the top-10 ranked genes used.