Abstract— Open source projects for example Eclipse and Firefox have open source bug repositories. User reports bugs to these repositories. Users of these repositories are usually non-technical and cannot assign correct class to these bugs. Triaging of bugs, to developer, to fix them is a tedious and time consuming task. Developers are usually expert in particular areas. For example, few developers are expert in GUI and others are in java functionality. Assigning a particular bug to relevant developer could save time and would help to maintain the interest level of developers by assigning bugs according to their interest. However, assigning right bug to right developer is quite difficult for tri-ager without knowing the actual class, the bug belongs to. In this paper, we have surveyed the enlisted reference papers for bug triaging in which the various triaged bug reports are assigned to the respective developers using data reduction techniques.
Key-Words: Text mining, classification, software repositories, open source software projects, triaging, feature extraction
1. Introduction
Data mining is the process of extracting useful information through data analysis. It is also known as knowledge discovery. Useful knowledge obtained as a result of data mining can be use to cut costs, increase revenues or both. Target data for mining purpose is categorical and numerical having data types like integer, decimal, float, char, varchar2 etc.
Data mining techniques cannot be applied to data that is not numerical or categorical. 85% of enterprise data falls in the category of non numerical or non categorical [1]. For the success of business, knowledge extraction from this unstructured data can be critical. Unstructured data is
processed using text mining techniques so that it can be processed by data mining algorithms and techniques. Techniques from information extraction, information retrieval and natural language processing are used by text mining .
Classification is a function of data mining to assign classes/categories to items in a collection. Basic goal of classification is the accurate prediction of target class for each case in data. For example, loan applications can be classified into high, medium or low risks on the basis of classification model.
1.1 Mining Software Repositories
To understand constantly evolving software systems is a very daunting task. Software systems have history of how they come to be and this history is maintained in software repositories. Software repositories are the artifacts that document the evolution of software systems. Software repositories often contain data from years of development of a software project [2].
Examples of software repositories are:
a) Runtime Repositories: Example of runtime repositories is deployment logs that contain useful information about application usage on deployment sites and its execution.
b) Historical Repositories: Examples of historical repositories are bug repositories, source code repositories and archived communication logs.
c) Code Repositories: Examples of code repositories are Google code and codeforge.net that store source code of various open source projects [3].
MSR is the process of software repositories analysis to discover meaningful and interesting information
hidden in these repositories. There is a huge Software Engineering data over the course of time. MSR picks this data, processes and analyzes it, and detects patterns in this data. MSR is an open field, both in what can be mined and what one can learn from the practice. Any software repository can be mined not necessarily the code, bug or archived communication repositories.
2. Problem Formulation
Triaging of bugs to developer to fix them is a tedious and time consuming task. Developers are usually expert in some particular area. For example few developers are expert in GUI and others are in pure java functionality. Assigning a particular bug to relevant developer could save time as well as would help to maintain the interest level of developers by assigning those bugs according to their interest. However assigning right bug to right developer is quite difficult for tri-ager without knowing the actual class a bug belongs to. This research proposes a technique for classification of open source software bugs using the summary provided by bug reporters.
2.1 literature survey
Some of the already implemented techniques for software bugs classification are:
a) Micheal W. Godfrey, Olga Baysal and Robin Cohen presented a framework for automatic assignment of bugs to developers for fixation using vector space model [4].
In this paper[4], the authors have proposed a specimen of the intelligent system that instinctively conducts the bug assignment. They have employed the vector space model to infer information about the developer’s expertise from the history of the previously fixed bugs. The vector model is used to retrieve the title and the description from the report to build a term vector which later can be used to find similar reports by mining the data in the bug repository. In order to create an efficient bug triage model, the authors conducted a survey wherein they collected a feedback from the developers regarding their previous bug fixing experience, their satisfaction with the bug assignment, whether they were successful and confident in handling bugs in the past, etc. The overall information provided them the initial estimates for the proposed model. This in turn helped them to implement the specimen model and test it within a software team working on the maintenance activities.
b) Hemant Josh, Chuanlei Zhouang, Oskum Bayrak presented a methodology to predict future bugs using history data [5].
In this paper[5], the authors have presented a bug prediction algorithm, the purpose of which is to predict the number of bugs which are to be detected and reported in each month. So the bug prediction of any month basically depends upon the bug count of the precursor month. All this prediction is achieved through the prediction alogorithm implemented in the respective paper[5].
c) Lei Xu, Lian Yu, Jingtao Zhao, Changzhu Kong, and HuiHui Zhang proposed a technique using data mining that automatically classifies the bugs of web-based applications by predicting their bug type.
In this paper[6], the authors have put forth the debug strategy association rules which find the relationship between bug types and bug fixing solutions. The debug strategy acquaints us with the erroneous part of the source code. Once the errors are found then it is very easy for the developers to fix them. The determined association rules help to predict files that usually change together such as functions or variables.
d) Nicholas Jalbert and Westley Weimer proposed a system that checks the duplicacy in the bug reports.
In this paper[7], the authors have instantiated a model that automatically indicates whether an arriving bug report is original or duplicate of an already existing report. It saves developer’s time. To predict bug duplication, system they have made use of rudimentary methods such as textual semantics, graph clustering and surface features.
e) Tilmann Bruckhaus provided a technique for Escalation Prediction to avoid escalations by predicting the defects that have high escalation risk and then by resolving them proactively [8].
3. Problem Solution
This section describes the proposed system for bug classification, data used for classification task and results obtained in different experiments.
3.1 Input Data
Eclipse and Mozilla firefox data is obtained from bugzilla -an open bug repository [9] [10]. Dataset of almost 29,000 record set is obtained. This data is divided into training and testing groups and experiments are performed on different set of data from these groups.
3.2 Model for prediction
When the bug is first reported to repository, it is submitted to our proposed system as shown in Fig. 1. System extracts all the terms in these reports using bag of words approach. The vocabulary is that of extremely high dimensionality and thus numbers of features are reduced by using feature selection algorithm. These features are used for training of classification algorithm which is then used for classification of bug reports. The classification algorithm used in proposed system is multinomial Naïve Bayes.
3.2.1 Pre-processing
Data pre-processing is the most important step of data mining. Data obtained from bug repositories is in raw form and cannot be directly used for training the classification algorithm. The data is first pre-processed to make it useful for training purpose. Data pre-processing is the major time consuming step of data mining and most important as well. Stop-words dictionary and regular expression rules are used to filter useless words and filter the punctuations respectively. Porter stemming algorithm is used to stem the vocabulary
3.2.2 Feature Selection
The vocabulary obtained after applying “bag of words” approach on data has very large dimensionality. Most of these dimensions are not related to text categorization and thus result in reducing the performance of the classifier. To decrease the dimensionality, the process of feature selection is used which takes the best k terms out of the whole vocabulary which contribute to accuracy and efficiency.
There are a number of feature selection techniques such as Chi-Square Testing, Information Gain (IG), Term
Frequency Inverse Document Frequency (TFIDF), and Document Frequency (DF) . In this research, we are using feature selection algorithm.
3.2.3 Instance Selection
The vocabulary obtained after applying “bag of words” approach on data has very large dimensionality. Most of these dimensions are related to Our pre defined bugs and thus result in reducing the performance of the classifier. To decrease the time, the process of Instance selection is used which takes the best k terms out of the whole vocabulary which contribute to accuracy and efficiency. This selection is fast instance of feature selection.
3.2.4 Classifier Modeling
Text classification is an automated process of finding some metadata about a document. Text classification is used in various areas like document indexing by suggesting its categories in a content management system, spam filtering, automatically sorting help desk requests etc.
Naïve Bayes text classifier is used in this research
for bug classification. Naïve Bayes classifier is based on
Bayes’ theorem with independent assumption and is a probabilistic classifier. INDEPENDENCE means the classifier assumes that any feature of a class is unrelated to the presence or absence of any other feature.
5. Conclusion
In open source bug repositories, bugs are reported by users. Triaging of these bugs is a tedious and time consuming task. If some proper class is assigned to these bugs it would be easier to assign these bugs to relevant developers to fix them. However, as reporters of these bugs are mostly non-technical it would not be possible for them to assign correct class to these bugs. In this research an automated system for classifying software bugs is devised, using multinomial Naïve Bayes text classifier. Feature selection algorithm and instant selection algorithm are used for bug triage. Maximum prediction accuracy is obtained.
6. Future Work
The main challenge to address in the future is strategic developers. In time, developers could learn how the system
assigns bug fixing tasks and try to manipulate task assignment. Thus, we should ensure that the assignment of bugs is a fair and manipulation-free process.
References
[1] A. Hotho, A. Nürnberger and G. Paaß, “A Brief Survey of Text Mining,” vol. 20, GLDV Journal for Computational Linguistics and Language Technology, 2005, pp. 19-62.
[2] A. E. Hassan, “The Road Ahead for Mining Software Repositories,”
IEEE Computer society, pp. 48-57, 2008.
[3] S. Diehl, H. C. Gall and A. E. Hassan, “Special issue on mining software repositories,” in Empirical Software Engineering An International Journal © Springer Science+Business Media, 2009.
[4] O. B. Michael and G. C. Robin, “A Bug You Like: A Framework for Automated Assignment of Bugs.,” IEEE 17th international conference, 2009.
[5] C. Zhang, H. Joshi, S. Ramaswamy and C. Bayrak, “A Dynamic Approach to Software Bug Estimation,” in SpringerLink, 2008.
[6] L. Yu, C. Kong, L. Xu, J. Zhao and H. Zhang, “Mining Bug Classifier and Debug Strategy Association Rules for Web-Based Applications,” in 08 Proceedings of the 4th international conference on Advanced Data Mining and Applications , 2008.
[7] N. Jalbert and W. Weimer, “Automated Duplicate Detection for Bug Tracking Systems,” in IEEE computer society, 2008
[8] T. Bruckhaus, C. X. Ling, N. H. Madhavji and S. Sheng, “Software Escalation Prediction with Data Mining,” in Data Mining, Fifth IEEE International Conference, 2006.
[9] [Online]. Available: https://bugzilla.mozilla.org/.
[10] [Online]. Available: https://bugs.eclipse.org/bugs/.