Discussion of “Application of Support Vector Machines in Assessing Conceptual Cost Estimates” by Sung-Hoon An, U-Yeol Park, Kyung-In Kang, Moon-Young Cho, and Hun-Hee Cho

Abstract

Support vector machines SVMs have been used in several areas of civil engineering since 1998 Gualtieri and Cromp 1998 . In the present study, the authors have employed SVMs for assessing the quality of conceptual cost estimates, which may be helpful in measuring the accuracy of the estimates. The study presented by the authors demonstrates the effectiveness of the SVM-based classification approach in comparison to the discriminant analysis approach. The discussed paper represents an important step toward an ongoing research effort for the use of SVMs for efficient modeling of the complex physical processes involved in solving various civil engineering problems. The objective of this discussion is to make some observations about issues related to the design of SVMs and the way accuracy assessment was carried out in this paper. The discussers feel that certain points require further clarification, which would help readers to understand the authors’ work. A number of research papers, such as Burges 1998 , referred to in the present study, discuss the theory of SVMs. In spite of this, the term “training datum” is used in place of “training data.” Per the discussers’ view, no such term as training datum exists in the design of SVMs. This requires a clarification from the authors. Further, for defining b bias term , a reference to Fig. 3 in the discussed paper is provided, which does not give any information about this parameter. In their discussion of the nonlinear support vector classifier, the authors mention “a kernel function satisfying this condition is known as the function satisfying Mercer’s condition.” We question the accuracy of this statement. A number of kernel functions are available in the literature, but only those kernels satisfying Mercer’s condition Vapnik 1999 are used as a kernel function in the design of SVMs. So the authors should clarify this paragraph. In the section on SVM for assessing conceptual cost estimates in the discussed paper, the authors suggest that “the SVM model is better suited to two-class problems than to multiple-class problems.” Again, the authors seem to be unclear about the theory of SVMs. Actually, the support vector classifier was designed for two-class problems Vapnik 1995 , which can easily be extended to multiclass problems by using an appropriate method. Per the discussers’ view, SVMs work very well for multiclass classifica-