@TECHREPORT{R-96-30,
NUMBER = {R-96-30},
INSTITUTION = ida,
ADDRESS = idaaddr,
YEAR = {1996},
AUTHOR = {Ohlsson, Niclas and Zhao, Ming and Helander, Mary},
TITLE = {Application of Multivariate Analysis for Software Fault Prediction
},
ABSTRACTURL = {/publications/cgi-bin/tr-fetch.pl?r-96-30+abstr},
ABSTRACT = {The need for quantitative methods to support project control has been expressed in a number of recent papers. A number of multivariate analysis techniques are available for analysing high­dimensional observations of software design metrics. This paper presents a successful study in which principal component analysis (PCA) and discriminant coordinates (DC) were used to develop prediction models for data from Ericsson Telecom AB. Instead of dividing modules into fault­prone and non­fault­prone, which has been common in previous studies, observations were categorised into several groups according to the ordered number of faults. The DC analysis revealed that the first discriminant coordinates statistically increase with the ordering of modules. This empirical result suggests an approach for ordering as a first step toward prediction of fault-prone modules that incorporates attributes of process and resources. The result of applying DC was compared with discriminant analysis (DA), which has been reported useful for building prediction models of fault-prone modules. The later models were found to be inadequate for predicting the most fault­prone modules for the considered data set. The authors experienced a number of problems while applying the earlier reported prediction models. These are illustrated in this paper, and improvements are suggested.
}