SOFTWARE DEFECT PREDICTION USING SOFTWARE METRICS WITH NAÏVE BAYES AND RULE MINING ASSOCIATION METHODS

• Main Author: Maruli Tua Simanguns, Fernando
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/36876
• Institution: Institut Teknologi Bandung

Producing software that does not contain defects or a little defects is not an easy task for software developers. Software testing is an important process to ensure software quality. Predicting software damage can help testers decide rational allocation of resources because they can find defects effectively, so they can improve software quality. Software defect prediction (SDP) is one solution for developers to achieve this. One that can be done in predicting defects is by using software metrics. Naive Bayes (NB) is one of the most used classification algorithms because of the simplicity of the algorithm and easy to implement, besides that NB also has very good performance and includes methods that require fast time for computation, and NB in classifying targets carries the interrelationship between attribute that builds a dataset because each attribute is considered independent. This paper supports the selection of features with the Association Rule Mining (ARM) method to find relationships between attributes associated with datasets that support predicting defects. The purpose of this study is to add the process of selecting features using ARM in the software prediction process using the NB method in the hope that it can improve accuracy of the method using software metrics. Software metrics have an association with one another in completing software, so this cannot be ignored. Three scenarios carried out to replace the proposed method: 1) Compare the proposed method (NB-ARM) with the NB method without feature selection; 2) Comparing the proposed method (NB-ARM) with the NB method with the selection of conventional features; and 3) Comparing the methods proposed by the SVM, NN and DTREE methods. By using the NASA MDP dataset which has 21 attributes (software metrics) and 1(one) target class with 2 (two) levels, namely defects and non-defects, empirical results from three scenarios in scenario 1(one) for parmeter precision, recall, f-measure, and improving increases accuracy are 0.101, 0.190, 0.154 and 0.180, and scenario 2 also increases by 0.106, 0.182, 0.159 and 0.163, as well as in the research scheme 3 also methods that show good performance using SVM, NN and Dtree with an average performance of 0.960 on the suggested method while the others are 0.855, 0.859 and 0.861. From the empirical results of the three scenarios created, the proposed performance method is better than the comparison method.