Journal of Software Engineering and Applications, 2011, 4, 639-645
doi:10.4236/jsea.2011.411075 Published Online November 2011 (
Copyright © 2011 SciRes. JS EA
Defect Prediction Leads to High Quality Product
Naheed Azeem, Shazia Usmani
Department of Computer Science, Feder al Urdu University of Arts, Science and Tech nology, Sindh, Pakistan.
Email: {naheedazeem, shaziausmani}
Received August 12th, 2011; revised September 25th, 2011; accepted November 5th, 2011.
Defect prediction is relatively a new research area of software quality assurance. A project team always aims to pro-
duce a quality product with zero or few defects. Quality of a product is correlated with the number of defects as well as
it is limited by time and by money. So, defect prediction is very important in the field of software quality and software
reliability. This paper gives you a vivid description about software defect prediction . It describes the key areas of soft-
ware defect prediction practice, and highlights some key open issues for the future.
Keywords: Defect Prediction, Software Qu ality
1. Introduction
Software life cycle is a human activity, so it is impossible
to prevent the injection of defects but it is possible to
produce the software with few defects. To deliver a de-
fect free software it is imperative to predict and fix the
defects as many as possible before the product delivers to
the customer.
Finding and fixing the defects after delivery usually
consumes a large portion of the project budget. Therefore,
defect prediction before delivery can contribute si gnify-
cantly to the success of project in terms of quality and
The aim of this research is to explore the different is-
sues and problems in the area of defect prediction as well
as provide the solutions to improve the product quality
via defect prediction mechanism.
In this survey report several research issues, formu-
lated as questions, need to be addressed to understand the
problems of defect prediction mechanism.
Research questions:
- How machine learning algorithms and data mining
techniques can be prove more effective in defect
extraction from repository?
- What kinds of software metrics are good indicators
of defects?
- To what extent the number of defects injected in the
software product can be reduced?
- How can we easily identify and localized the soft-
ware defects?
- What kind of software repository could represent
the required information?
- What methods or procedures are better to opt for
defect identification and localization?
- How can we reduce the probability of false alarm?
- How good predictor is in finding actual defective
modul es ?
- Is there a type of defect prediction model that pro-
vides a good fit to defect-prediction across multiple
releases and in many organizations?
- To what extent can we use other project data to pre-
dict defects for a software system and is there any
possibility to transfer prediction models from one
project to another?
The rest of this paper is outlined a s follows. We begin
by providing background and descriptions in section 2.
The Issues and problems encountered by the defect pre-
diction are elaborated in section 3. The methods and ap-
proaches used to tackle issues are illustrated in section
4.In section 5 some future research areas are presented.
Finall y, we fin i s h wit h conc l u sio n.
2. Background and Descriptions
A software defect is an error, flaw, mistake, failure, or
fault in a computer program or system that produces an
incorrect or unexpected result, or causes it to behave in
unintend ed ways.
Software defects are expensive in ter ms of qualit y and
cost. Moreover, the cost of capturing and correcting de-
fects is one of the most expensive software development
activities. It will not be possible to eliminate all defects
but it is possible to minimize the number of defects and
their severe impact on the projects. To do this a defect
Defect Predict ion Leads to High Quality Product
Copyright © 2011 SciRes. JS EA
management process needs to be implemented that fo-
cuses on improving software quality via decreasing the
defect density. A little investment in defect management
process can yield significant returns.
Software De fect Pred ictio n
The most discussed problem is software defect prediction
in the field of software quality and software reliability.
As Boehm observed finding and fixing a problem after
delivery is 100 times more expensive than fixing it dur-
ing requirement and design phase. Additionally software
projects spend 40 to 50 percent of their efforts in avoid-
able rework.
In summarizing the major research trends for defect
prediction of software products include:
Software Met ri cs
Software metric is a measure of some characteristic or
attribute of software module. Since Software metrics are
quantitative methods and have proved so powerful in
defect prediction. The essence of software quality engi-
neering is to investigate the relationship between differ-
ent metrics and end -product quality.
Software metrics can be classified into three categories:
product metrics, process metrics, and project metrics.
Product metrics describe the characteristics of the prod-
uct such as size, complexity, design features, performan-
ce, and quality level. Process metrics can be used to im-
prove software development and maintenance. The pro-
ject parameters such as the number of developers and
their skill le vels, the schedule, the size, a nd t he organiza-
tion structure certainly affect the quality of the product.
Defect Identif ic ation
Identifying a nd locating d efects in soft ware proj ects is
a difficult task. Further, estimating the density of defects
is more difficult. So, the software project team is fully
focused on finding and fixing all the defects.
Defective and Defect-free modules
Accuracy of defect prediction techniques is determi-
ned b y corr ectly find ing the defective parts of a software
pro d uct witho ut gi vi n g an y false a la r m. G i vi ng hi gh fa l se
alarm rate means developers and testers wasting their
time in inspecting and testing defect free modules. On
the other hand, predictions of defective modules as defect
free modules would cause more expensive in terms of
quality and co st.
Data Mi ning and Machine Learning Techniques
Machine learning models and Data mining techniques
can be applied on the software repositories to extract the
defects of a software product. Common algorithms inclu-
de decision tree learning, Naive Bayesian classification
and ne ural networks.
3. Issues and Problems
3.1. Problem with Selecting the Right Set of
Studies based on accuracy of defect prediction
model focused on either project metrics or product
metrics but not the combined impact of both [1].
It is strongly believed that software size has a re-
lationship with software quali ty but there is a lac k
of evidence that shows size metrics as a good in-
dicator of defects [2].
Defect prediction model did not conclude that ei-
ther change metrics or code metrics were better
for defect removal [3].
Managers rely on complexity metrics to allocate
QA resources effectively, but complexity metrics
fail to predict critical binaries of a complex sys-
tem [4].
The effectiveness of SSM as defect predictor in
OO software needs to be established [5].
3.2. Problem in Red u cing Fa lse Alar m
It is difficu lt to find the o ptimum th reshol d v alu e that
makes the difference between defective and non de-
fective modules [6].
Static attribute s are mainly u sed in decreasing fal-
se alarms but they do not provide the enough in-
formation to significantly reduce the rate of false
alarms [7].
The influence of refactoring on defects prediction
process do not provide any conclusive result i.e.
either refactoring or non refactoring related features
leads t o high quali ty def e ct predict i on m odel [8] .
Implementation of Defect management process in
multi-site software development organization is
difficult and more challenging [9].
Relationship between change coupling and soft-
ware defect were unknown [10].
3.3. Defect Ident i fica tion Issue
Most of the defect pred ic tion models d o not utiliz e
customer profile and system characteristics to p re-
dict the customer reported defects in order to im-
prove the defect prediction mechanism [11].
Traditional capture recapture model do not esti-
mate the number of defects in post inspection ph-
ase and also rely on expert inspectors [12].
Simulation approach and queuing theory used to
model defect removal process did not consider the
utilization of developers [13].
Only the number of defects cannot provide enou-
gh information to support the software quality ac-
tivities [14 ] .
Defect Predict io n Leads to H igh Quality Prod uct
Copyright © 2011 SciRes. JSEA
Traditionally, some defect prediction models are
used to identify the number of defects in a multi-
version system but they are not platform and
language independent [15].
3.4. Problem with Extraction of Defects from
Extraction of defects from software bug repository
accurately is not done without a good data mining
model [16].
There is a need of good data mining model to pre-
dict the software defects from a bug repository
Most of the machines learning algorithms are not
capable of extracting defects from the database
that store continuous features [18].
Prior research on defect prediction fails to fully
utilize the defect data and defect repair time esti-
mation requires mathematical a ssumptions [19].
Software prediction model only works well when
enough amount of data is available in software re-
pository within the orga nizati on to init iall y fed t he
model [20].
4. Approaches and Methodologies
4.1. Metrics as a Predictor of Software Defects
Only the few authors claimed that project metrics are
also helpful to improve the quality of software product.
Wahyudin and Schatten studied the relationship of
project metrics to the potential growth of defects and
determined the combined impact of project metrics and
product metrics on defect prediction .Two step predictor
pro- cess was proposed in [1]. First find out the Pearson
rank correlation among the strongest correlated predic-
tors with the dependent predictor (variable) and then use
the stepwise linear regression and backward elimination
to exclude the ins ignificant predicto rs to form a reliabili-
ty gro wth mode l. Result revealed that project metrics has
strong correlation with the potential growth of defects
between release and combined effect of project and
prod uct metric s were result in better pr e diction model.
Line of code (LOC) is one of the simplest and widely
used metric. H. Zhang analyzed the. Two public defect
datasets to prove the relationship of software size with
software quality. The “ranking ability” of LOC can be
actually modeled by a Weibull distribution function. By
using defect density values calculated from a small per-
centage of the largest modules, LOC’s ability to predict
the number of defects can be improved. Also using typi-
cal classification techniques, defective components based
on LOC are able to predict. Results showed that LOC can
be a useful indicator of software quality, and useful to
build defect prediction models using LOC [2].
In [3] another comparative analysis was done to check
the effectiveness of change metrics over code metrics for
defect prediction. Three different models one for change
metrics, one for code metrics and one for both change
and code metrics, using three machine algorithms i.e.
logistic regression, Naive Bayes and decision tree (J48).
Cost analysis was also performed to evaluate the cost
associated with the prediction errors of these three mod-
els. Result showed that change metrics are significantly
better indicator for defect prediction model than static
code attr ibutes.
Zimmermann studied that complexity metrics are
failed to pro vide better result i n defect predic tion when i t
comes to critical binaries. A dependency graph of win-
dows 2003 is build. For each node (binary) on dependen-
cy graph, network measures are computed. Several code
metrics as a co ntrol set are applied to all the b inaries [4].
Fenton et al. criticized that there is no relationship
between complexity and defects as well as with the size.
In spite of critique, most of the studies used SSM as in-
dicator of defects in procedural paradigm as well as in
OO software. The role of software science metrics (SSM)
in defect prediction of object oriented (OO) software had
been studied. Binary and numeric classification models
available in WEKA are applied on dataset with class lev-
el data. T he models are first applied using a ll the metrics
available in the dataset and t hen re movin g SSM fro m the
input and the accuracies and error values of all the mod-
els are observed. Effectiveness of SSM is measured at
model level by comparing accuracies and Mean absolute
error of models with and without SSM [5].
4.2. Software Defect Reduction
Performance of defect prediction mechanism is deter-
mined by the probability of defect detection and proba-
bility of false alarms. Reduction in the false alarm is do-
ne by a two-dimensional ROC analysis. The author cha-
nged the decision threshold on Naïve Bayes and obser-
ved the changes in prediction performance measures.
Using decision threshold optimization on Naïve Bayes
classifier, probability of false alarm (pf) rate has decrea-
sed, while Balance rate has increased and the probability
of detection (pd) rates remained the same. These results
were also validated using paired t-test [6]. 10 repository
metrics were extracted from CVS revision system o f ec-
lipse project and classified the source files as defective if
it contained one or more defects and non defected if it
contained zero defect. Naïve Bayes used to extract the
additional metrics from repository to increase the input
data. Mann-Whitney U test was done on data to test the
statical significance of us ing differe nt technique s. Results
showed that repository metrics give better insight to
software product and hence able to lower pf rate as com-
pared to using only static cod e a ttributes [7].
Defect Predict ion Leads to High Quality Product
Copyright © 2011 SciRes. JS EA
Earlier studies have addressed evolutionary activities
areas such as refactoring based on history information. It
was also investigated the influence of refactoring on de-
fects prediction process. Approach discussed in [8] pro-
vide c oncl usive r esult that refac toring l ead s to hi gh qua l-
ity defect prediction model. Authors identify the refac-
toring features using evolution data extracted from ver-
sioning system. Machine learning algorithm WEKA is
used to generate the section model made up of same
number of defected and non defected files, and then the
statistical analysis is performed on these models to eva-
luate the hypothesis.
In a multi-site organization when development is dis-
tributed, hundreds of people are working on a project;
establish a defect prediction mechanism is too difficult.
A GQM (Goal Question metric) method is adopted and a
combination of three quality metrics was used [9].
Ambros and Lanza studied the relationship between
coupling and software defects. They create a source code
model and a history model and train them with the nec-
essary data. The correlation of change coupling with the
number of software defects, major defects and with se-
vere defects is measured by using Spearman correlation
coefficient on three large java systems [10] as shown in
Figure 1.
4.3. Identifying and Locating Defects
Most of the researchers have focused on predicting the
number of defects to improve the defect prediction mech-
anism but only the numbers of defects is not enough to
get s ufficient information that support quality activities.
Raaschou and Rainer raise this issue and build a model
that based on customer reported defects. Exposure model
takes detailed customer profile, reported defects and ver-
sion data as input. Number of new defects at low level
can be expressed as the product of the defected versions
at initial release and isolated effect of five exposure fac-
Figure 1. Creating a model with bug and change coupling
information [10].
tors. These factors influence the number of defects found
and can be aggregated as needed to higher levels [11].
Bucholz investigated that static capture recapture mo-
del did not capture the defects when it is in post release
phase. They extend the static capture recapture model. In
dynamic capture recapture model for initial release, 30-
40 size of capture recapture is fixed and for the next re-
lease heuristic algorithm is used to calculate the next
capture recapture size. Identify the number of defects in
each release and find out the duplicates by matching new
reported defects to the previous defects existing in defect
database. Calculate the total number of predicted defects
by using Peterson estimator [12]. Fan and Xiaohu consi-
dered the developer data. Defect detection process was
done by an algorithm which generates non homogenous
poisonous process. Classified the new defects according
to their seve rity a nd the n assi gned to the deve lop ers. U p-
date the status of defects and fixed defects were removed
from the q ueue [13].
Hong and Baik proposed a new approach for predict-
ing the distribution of defects and their types based on
project characteristics as can be seen in Figure 3. D ete r-
mine the factors that affect software attributes being es-
timated then perform behavior analysis on them. Gather
all the defect data and build a model. The model for de-
fect prediction was built by using curve fitting method
and regression analysis. After statistical modeling and
regression analysis validates and refines the model. [14].
Kastro build a model that worked on different version
of software and it is language and platform independent
(Figure 2). In [15] Metric data include CVS level data,
Figure 2 . Proposed defect predi ct ion model [15].
Defect Predict io n Leads to H igh Quality Prod uct
Copyright © 2011 SciRes. JSEA
Figure 3 .Overview of t he pr oposed model [14].
change level data and previous version data was collec -
ted. Collected data was organized, consolidate and nor-
malized. Built a model by using multilayer perception
with neural networks and trained the model with norma-
lized data.
4.4. Data Mining and Software Repositories
Software rep ositor ies have lo ts o f infor mation t hat is use -
ful in a ssessin g soft ware quali ty. Data mining techniques
and machine learning algorithms can be applied on these
reposito ries to extract the usef ul in formation.
Tosun and Bener applied AI technique in predicting
defects. They extract static code attributes at functional
level from the source code and then store the de fect data.
They construct and calibrate the defect prediction model
using AI algorithm [16]. In [17] a two step data mining
model is proposed to predict software bug estimation. In
first step, a weighted similarity model is used to match
the summary and description of new bug from the pre-
vious bug in the bug repository. In the second step cal-
culate the duration of all the bugs and the average is cal-
The technique used in [18] is entropy based splitting
criteria and minimum description stopping criteria (de-
cide when to stop discretization).The binary discretiza-
tion was always selecting the best cut point and was ap-
plied recursively (Table 1). The authors investigated the
effect of discretiza tion on defect pred ic tion models.
Hewett proposed a model that aid help in software test-
ing and estimate defect repair time. Empirical approach
employed data mining technique that increased the utili-
zation of defect data in prediction of defect repair time to
support testing and defect management. They used four
Table 1. Summary of the res ults [1 8].
Classifier Software Modules
Corre ctly Cl a ss if ied Instances
Dis cretization % After
Discretization %
Naïve Bayes 121 85.124 95.562
101 82.178 85.148
J48 121 90.082 92.562
101 83.168 85.148
data mining algorithms based on three different appro-
aches i.e. decision tree learner, Naïve Bayes classifier
and neural network approach used to build a mode [19].
Zimmermann findings are in the case of when there is
not enough historical data to train the model. In that sce-
nario Zimmermann used cross project data to build and
train the model. Relative measures such as code churn
(added, deleted, and changed lines), domain metrics and
process metrics extracted from development process of
one project are used to build prediction model for another
project, based on logistic regression. Their research result
as can be seen in Figure 4. Accuracy, precision and re-
call are used to assess the model [20].
5. Future Work and Open Issues
Futur e work in this area should:
- Establish an improved method for predicting soft-
ware quality via identifying the defect density of fa-
ult prone modules and improve the rate of false al-
arm [6,8,12].
- Different machine learning algorithm and data min-
ing techniques are used to improve the defect pre-
diction accuracy [16,18].
- Extracting automatically key information from the
data repositories that are more relevant for defect
predictio n [7] and trying right set o f metrics that in-
fluence the success of cross project predictions [19,
6. Conclusions
Software defect prediction is the process of locating de-
fective modules in software. To produce high quality
software, the final product should have as few defects as
possible. Early detection of software defects could lead
to reduced development costs and rework effort and mo-
re reliable software. So, the study of the defect prediction
is important to achieve software qualit y.
Figure 4. Results from 622 cross-project defect predictions.
For example, Firefox data can predict IE. The color tells
whether a project predicts other projects (white), can be
predicted (bl a ck), or both (gray) [20].
Defect Predict ion Leads to High Quality Product
Copyright © 2011 SciRes. JS EA
Therefore our study aims to provide useful insight on
the defect prediction approaches to aid project team in
making quality product.
The findings of this research confirm observations made
by other researchers that:
- History metrics extracted from repository helps in
reducing false alarm as well as increasing the rate of
probability of detection in open source software and
LOC has positive relationship with software de-
- Refactoring has great impact on software quality
improvement as well as on building high quality
defect prediction model.
- GQM method with combinations of quality metrics
proved better to support the defect prediction proc-
ess in multi-site organization s.
- The number of defects alone cannot be sufficient
information to provide the basis for planning quality
assurance activities and assessing them during ex-
ecution. That is, for project management to be im-
proved, we need to predict other possible informa-
tion about software quality such as in-process de-
fects, their types, and developer and customer pro-
files and so on.
- Change coupling is correlated with the number of
software defects, major defects and with severe de-
- Integration of discretization method with classifica-
tion algorithm improves the defect prediction accu-
racy by transforming the continuous features into
discrete features.
- Data mining techniques are use ful in prediction of de-
fec t repair time, software bug estimation more accu-
rately, and predicting the number of defects in mult i-
version environment that is language and platform
- Cross project defect predictors build accurate pre-
diction model when signi fican t factor s are evaluated
and quantified.
- Network measures on dependency graph predict de-
fect for critical binaries more accurately than com-
plexity metrics.
- Use of software science metrics is ineffective for
defect prediction and classification of defect prone
modules in object oriented software.
- The comparative study on product metrics and pro-
cess metrics concluded that overall change metrics
were effectively better than code metrics as well as
project metr ics with co mbination of pr oduct metrics
is more effective in defect prediction.
[1] D. Wahyudin, A. Schatten, D. Winkl er , A. M. Tjoa and S.
Biffl, “Defect Prediction Using Combined Product and
Project Metrics: A Case Study from the Open Source
“Apache” MyFaces Project Family,” Proceedings of the
2008 34th Euromicro Conference Software Engineering
and Advanced Applications, 2008, pp. 207-215.
[2] H. Zhang, “An Investigation of the Relationships between
Lines of Code and Defects,” 2009 IEEE International
Conference on Software Maintenance, pp. 274-283.
doi:10.1109 /I C S M. 20 09.53 06 304
[3] R. Moser, W. Pedrycz and G. Succi, “A Comparative
Analysis of the Efficiency of Change Metrics and Static
Code Attributes for Defect Prediction,” In Proceedings of
the International Conference on Software Engineering
(ICSE’08 ), Leipzig, pp. 181-190.
[4] T. Zimmermann and N. Nagappan, “Predicting Defects
using Network Analysis on Dependency Graphs,” In
Proceedings of the International Conference on Software
Engineering (ICSE’08), Leipzig, Germany, pp. 531-540.
[5] Z. A. Rana, S. Shamail and M. M. Awais, “Ineffective-
ness of Use of Software Science Metrics as Predictors of
Defects in Ob ject Orien ted Software,” Proceedings of the
2009 WRI World Congress on Software Engineering, Vol.
04, 2009, pp. 3-7. doi : 10.1109/WCSE.2009.92
[6] A. Tosun and A. Bener, “Reducing False Alarms in
Software Defect Prediction by Decision Threshold Opti-
mization,” Third International Symposiumm on Empirical
Software Engineering and Measurement, 2009 IEEE, pp.
477-480. doi:10.1109 / E S EM . 20 09.53 16 006
[7] B. Cagl ayan, A. Bener and S. Koch, “Merits of Using
Repository Metrics in Defect Prediction for Open Source
Proj ects,” 2009 ICSE Workshop on Emerging Trends in
Free/Libre/Open Source Software Research and Devel-
opment, pp. 31-36.
[8] J. Ratzinger, T. Sigmund and H. C. Gall, “On the Rela-
tion of Refactoring and Software Defects,” In Proceed-
ings of the International Workshop on Mining Software
Repositories (MSR ’08), Leipzig, Germany, pp. 35-38.
[9] K. Korhonen and O. Salo,Exploring Quality Metrics to
Support Defect Management Process in a Multi-Site Or-
ganization—A Case Study,” 19th International Sympo-
sium on Software Reliability Engineering(IEEE '08), pp.
[10] M. D’Ambros, M. Lanza and R. Robbes,On the Rela-
tionship between Change Coupling and Software De-
fec ts,” Proceedings of the 2009 16t h Worki ng Con ference
on R evers e Engineer ing, IEEE 2009, pp.135-144.
[11] K. Raaschou and A. Rainer, “Exposure Model for Predic-
tion of Number of Customer Reported Defects,” Pro-
ceedings of the Second ACM-IEEE International Sympo-
sium on Empirical Software Engineering and Measure-
ment, (ACM ’08), pp. 306-308.
[12] R. Bucholz and P. A. Laplante,A Dynamic Cap-
ture-Recapture Model for Software Defect Prediction,”
Innovations in Systems and Software Engineering (2009),
Spring e r L on don, pp. 265-270.
[13] W. Fan, Y. Xiaohu, Z. Xiaochun and C. Lu, “Simulation
of the Defect Removal Process with Queuing Theory,”
Defect Predict io n Leads to H igh Quality Prod uct
Copyright © 2011 SciRes. JSEA
3rd International Symposium on Empirical Software En-
gineering and Measurement, 2009, pp. 473-476.
[14] Y. Hong, J. Baik, I. Y. Ko and H. J. Choi,A Value-
Added Predictive Defect Type Distribution Model based
on Project Characteristics,” Seventh IEEE/ACIS Inte rna-
tional Conference on Computer and Information Science,
2008, pp. 469-474.
[15] Y. Kastro and A. B. Bener, “ A Defect Prediction Method
for Software Versioning,” Software Quality Journa l ,
Spri nger Nether lands, Vol . 16, 2008, pp. 543-562.
[16] A. Tosun, B. Turhan and A. Bener, “Practical Con sidera-
tions in Deploying AI for Defect Prediction: A Case
Study within the Turkish Telecommunication Industry,”
Proceedings of the 5th International Conference on Pre-
dictor Models in Software Engineering, 20 09, pp. 24-25.
[17] N. K. Nagwani and S. Verma, “Predictive Data Mining
Model for Software Bug Estimation Using Average W-
eighted Similarity,” IEEE 2nd International Advance
Computing Conference, 2010, pp. 373-378.
[18] P. Singh and S. Verma, “An Investigation of the Effect of
Discretization on Defect Prediction Using Static Meas-
ures,” IEEE International Conference on Advances in
Computing, Control, and Telecommunication Technolo-
gies, 20 09 , pp. 837-839. doi:10.1109/ACT.2009.212
[19] R. Hewett, “Mining Software Defect Data to Support So-
ftware Testing Management,” Applied Intelligence, Sp-
ringe r Ne the r l a nds , 20 09.
[20] T. Zimmer mann, N. Nagappan, H. Gall, E. Giger and B.
Murphy, “Cross-Project Defect Prediction, a Large Scale
Experiment on Data vs. Domain vs. Process,” European
Software Engineering Conference and Symposium on the
Foundations of Software Engineering (ESEC/FSE),
Amsterdam, 2009, pp. 91-100.