Transcript Slides - Crest

Real-World Challenges in
Building Accurate Software
Fault Prediction Models
DR. ÇAĞATAY ÇATAL
TUBITAK (Research Council of TURKEY)
Predictive Modelling and Search Based Software Engineering, London, UK, 24-25 October 2011
Outline

Introduction



Challenging Issues








Dependable Software Systems
Motivation
Fault prediction with no fault data
Fault prediction with limited fault data
Noise detection on measurement datasets
Practical tools (Eclipse plug-in)
Cross company vs. Within-company fault prediction
Our Models
A Systematic Review Study
Conclusion
2
Dependable Systems

Are we successful in building dependable software
systems?
Safety
 Security
 Reliability
 Availability

(not being harmful for environment)
(ability to protect the privacy)
(ability to perform its function for a period of time)
(ability to serve whenever needed)
3
1. BRITISH ATM PAYS DOUBLE !
19 March 2008
Hull, England
• ATM pays out double the amount withdrawn
• Dozens of customers lined up in front of ATM
• This continued until ATM ran out of money at 8 p.m.
4
A Generous British ATM...
A Sainsburry’s spokesman said
“ We do not know how much the machine paid out at the
moment but the matter is under investigation”
A customer said
“ I joined the queue and when I finally got to the front I drew
out 200 pound but it gave me 400 pound. The statement said
I only drew out 200 pound. I don’t know whether I will have
to pay it back”
The police said
“ Those who benefited could face charges, but only if the
company administering the machine complained”.
5
2. ATM Pays Out Double the Cash, 16 January 2009
6
3. Tesco machine pays double,
18 August 2009
7
4. Dundee cash machine, 20 January 2011
8

But what happens if an ATM malfunctions and pays
out less than you asked for?

We need dependable systems !
9
Motivation

Project Managers ask several questions:
 “How can I get the code into production faster?
 What code should we refactor?
 How should I best assign my limited resources to
different projects?
 How do I know if code is getting better or worse
as time goes on?”
•Software Metrics
•Software Fault Prediction
“Baseline Code Analysis Using McCabe IQ”
10
Example: gcc project




/trunk/gcc/fold-const.c
http://gcc.gnu.org/viewcvs/trunk/gcc/foldconst.c?revision=135517&view=markup
fold_binary’s CC value is 1159 !
Security problems or faults can occur
11
Vulnerability Report – Fold_Binary Method
http://vulnerabilities.aspcode.net/14389/fold+binary+in+fold+const+c+in+GNU+Compiler+Col.aspx
12
CHAPTER 2:
Challenging Issues
Software Fault Prediction Modeling
Current
Project
Previous
Version
Training
Learnt
Hypothesis
Software
Metrics
Unknown Fault Data
Known Fault Data
Software
Metrics
Predict
Faults
14
1. No Fault Data
Software
Metrics
Previous
Version
Learnt
Hypothesis
Predict
Faults
Unknown Fault Data
Software
Metrics
Unknown Fault Data
Training
Current
Project
* How does the software quality assurance team predict the software
quality based on only the recorded software metrics?
- A new project type for organization
- No quality measurement have not been collected
* Supervised learning approach cannot be taken
15
2. Limited Fault Data
Known Fault Data
Software
Metrics
Previous
Version
Learnt
Hypothesis
Predict
Faults
Unknown Fault Data
Software
Metrics
Unknown Fault Data
Training
Current
Project
* During decentralized software development, some companies may not
collect fault data for their components
* Execution cost of data collection tools may be expensive
* Company may not collect fault data for a version due to the lack of budget
- Can we learn both from labeled and unlabeled data?
16
3. Noise Detection

Noisy modules degrades the performance of
machine learning based fault prediction models




Attribute Noise
Class Noise
Class noise impact classifiers more severely as
compared to attribute noise
We need to identify noisy modules if they exist
Some cases:
 Developers may not report the faults
 Data entry and data collection errors
17
4. Practical Tools












Earliest Work, Porter and Selby, 1990
....
Logistic Regression (Khoshgoftaar et al., 1999)
Decision Trees (Gokhale et al., 1997)
Neural Networks (Khoshgoftaar et al., 1995)
Fuzzy Logic (Xu, 2001)
Genetic Programming (Evett et al., 1998)
Case-Based Reasoning (Khoshgoftaar et al., 1997)
Pareto Classification (Ebert, 1996)
Discriminant Analysis (Ohlsson et al., 1998)
Naive Bayes (Menzies et al., 2008)
...
Hundreds of research papers but lacking of practical tools…
18
5. Cross-Project vs. Within-Company
Fault Prediction

Can we use cross-company (CC) data and predict
the fault-proneness of program modules in the
absence of fault labels?
19
CHAPTER 3:
Models we built...
1. No Fault Data
21
1. No Fault Data Problem- Literature

Zhong et al., 2004, Clustering and Expert based Approach
 K-means and Neural Gas algorithms
 Mean vector and several statistical data such as min., max.
 Dependent on the capability of the expert

Zhong, S., T. M. Khoshgoftaar, and N. Seliya, “Unsupervised Learning for Expert-based Software
Quality Estimation”, Proceedings of the 8th Intl. Symp. on High Assurance Systems Engineering,
Tampa, FL, 2004, pp. 149-155.
22
1. No Fault Data Problem
1. Our technique first applies X-means
clustering
method
to
cluster
modules and identifies the best
cluster number.
2. The mean vector of each cluster is
checked
against
the
metrics
thresholds vector. A cluster is
predicted as fault-prone if at least
one metric of the mean vector is
higher than the threshold value of
that metric.
[LOC, CC, UOp, UOpnd, TOp, TOpnd]
[65, 10, 25, 40, 125, 70]
(Integrated Software Metrics (ISM)
document)
23
• Datasets from Turkish white-goods
manufacturer
• Effective results are achieved
• No expert opinion
• Identification of threshold vector is
difficult
24
2. Limited Fault Data Problem
25
2. Limited Fault Data Problem

We simulated small labeled-large unlabeled data problem with
5%, 10%, and 20% rates and evaluated the performance of
each classifier under these circumstances.

Naive Bayes algorithm, even if it is a supervised learning
approach, works best for small datasets

YATSI (Yet Another Two Stage Idea) improves the
performance of Naive Bayes algorithm for large datasets if the
dataset does not consist of noisy modules

We suggest Naive Bayes for limited fault data problem as well
26
3. Noise Detection
27
3. Noise Detection
Our hypothesis:

A data object that has a non-faulty class label is
considered a noisy instance if the majority of the software
metric values exceed their corresponding threshold values.

A data object that has a faulty class label is considered a noisy
instance if all of the metric values are below their
corresponding threshold values.

How to calculate software metrics threshold values?
R. Shatnawi, W. Li, J. Swain, T. Newman, Finding software metrics
threshold values using ROC curves, Journal of Software Maintenance and
Evolution: Research and Practice 22 (1) (2010) 1–16.
28
How to Calculate Threshold Values



The interval for the candidate threshold values is between the
minimum and maximum value of that metric in the dataset.
Shatnawi et al. (2010) stated that they chose the candidate
threshold value that has the maximum value for both
sensitivity and specificity, but such a candidate threshold
may not always exist.
We calculated the AUC of the ROC curve that passes through
three points, i.e., (0, 0), (1, 1), and (PD, PF), and we chose the
threshold value that maximizes the AUC.
29
30
31
4. Practical Tools
32
4. Eclipse based Plug-in (RUBY)
33
Sample User Interfaces - Features
34
Result Views
35
5. Cross-Project Fault Prediction
36
5. Cross-Project Fault Prediction

We developed models based on software metrics
threshold values



If majority of software metrics thresholds values are
exceeded, the label of the module is faulty
Otherwise, non-faulty label is assigned
Threshold values are calculated from the other
projects (cross-company)
37
38
AUC, PD, PF
39
Results




Case studies showed that the use of cross-company data is
useful for building fault predictors in the absence of fault
labels and remarkable results are achieved.
Our threshold-based fault prediction technique achieved
larger PD (but larger PF) value than Naive Bayes based
approach.
For mission critical applications, PD values are more
important than PF values because all of the faults should be
removed before deployment.
In summary, we showed that cross-company dataset is useful.
40
4. Systematic Review
41
A Systematic Review Study

74 papers published between 1990 and 2007
27 journal papers
 47 conference papers


We report distributions before and after 2005, since that was
the year that the PROMISE repository was established.
42
Results

The journals that published more than two fault model
papers are: IEEE Transaction of Software Engineering (9);
Software Quality Journal (4); Journal of Systems and Software
(3); Empirical Software Engineering (3)

14% of papers were published before 2000 and 86% after.

Types of data sets used by authors were: private (60%),
partial (8%), public (31%), unknown (1%). ‘‘Partial” means data
from open source projects that have not been circulated.

Since 2005 the proportion of private datasets has reduced to
31%, the proportion of public data sets has increased to 52%.
There are 14% partial datasets and 3% unknown.
43
Results (cont’d)

Data analysis methods are machine learning (59%), statistics
(22%), statistics and machine learning (18%) and statistics and
expert opinion (1%).

After 2005 the distribution of methods is machine learning
(66%), statistics (14%), statistics and machine learning (17%)
and statistics and expert opinion (3%).

60% of papers used method level metrics, 24% used class
level metrics, 10% were file level metrics, other categories less
than 5%. 2005, 53% were method level, 24% were class level
and 17% were file level (others less than 3%).
44
Suggestions



More studies should use class-level metrics to support
early prediction.
Fault studies should use public datasets to ensure results
can be repeatable and verifiable.
Researchers should increase usage of machine learning
techniques.
45
Conclusion & Future Work




Software fault prediction is still challenging and quite useful
We need practical tools
Prediction models can be used to predict vulnerability-prone
modules
Challenges
 How to make fault prediction work across projects ?
 How to build models when there is no fault data?
 How to build models when there is very limited fault data?
 How to remove noisy modules from datasets?
46
THANK YOU
Cagatay CATAL, Ph.D.
[email protected]
www.cagataycatal.com
47