从 H V 序 聚 类 的 软 fl 缺 阽 m 测 力H JI /l
A b st ra c t
A b str a c t
Softw are Q uality A ssurance (SQ A ) is one of the m ost im portant steps of softw are
eng ineering, w h ich evaluates the quality of softw are prod ucts by using a series of
assurance m ethod s to assu re to deliver a perfect so ftw are pro d uct to the m anager and
custom ers. H ow ever, since the d ual pressure of tim e to m arket and co st contro l,
so ftw are dev elo pm ent m anagers often find it hard to allocate su fficient tim e and
reso urces fo r so ftw are test and other quality assurance activ ities. T h is alw ay s results
in th e defect-pro ne pro ducts to cu stom ers，w hich o ften leads to th e erro r in the
so ftw are run n in g and ev en cause serious pro blem s in system dependab ility and safety.
T here is so m e research w ork o n the softw are defect pred iction . So m e is to b uild
so ftw are defect pred ictio n m odels by em p irical study, w h ich has been proved as a
cost-effective w ay to allev iate th is prob lem . W ith the em p irical study on the m etrics
o f softw are artifact o r so ftw are develo pm ent process and the defect d ata, researchers
can bu ild the de fect p red ictio n m od els to p red ict th e d efects n um ber o r the likelih oo d
a softw are file o r class contain s defect. B y using the defect pred ictio n m odels, th e
d efects conta in ed in th e so ftw are p ro d ucts can be p red icted , w h ich ca n assist th e
softw are develo pers in the so ftw are q uality assu ran ce and testing m ore effectively.
T h e research o f th is paper fo cuses o n the so ftw are defect p red iction m o de ls,
especially the m ode ls based on the so ftw are p roduct m etrics. T he prior research in th is
d irection m ainly pred icts defects by class o r file, i.e. treat classes/files as basic u n its
for defect evaluatio n and rank ing, an d im p ro ves the perform ance of the pred ictio n
m o dels based on class/file. S om e w ork pro poses to pred ict defects o n the h igher
package level and gets better recall and p recision. H ow ever, it is also repo rted that
package-based pred iction m odels are less effectiv e than class-based p red ictio n m ode ls
a s fa r a s th e e ffo rt is c o n c e rn e d .
In th is paper, w e p ropo se a n o vel softw are defect pred ictio n m ethod based o n
fun ctional clusters o f program s. In the m ethod , w e use pro per-grained and
pro blem -oriented program clu sters as the basic un its of defect predictio n in order to
im pro ve the perfo rm ance, especially the effort-aw are perfo rm ance, of defect
pred ictio n.
T o evaluate the effectiveness o f the m eth od , w e conducted an ex perim enta l stu d y
on E clipse 3 .0 sy stem . W e em p loyed d ifferent d ata an aly sis m ethod s to bu ild the
IV
