Investigating Bad Smells with Feature Selection and Machine Learning Approaches

Code Smell is a piece of code that is designed and implemented poorly and it gives adverse effect on the software quality and maintenance. Now, a day’s machine learning based techniques have been extensively used towards code smell research. The main objective of this research is to optimise the features of Android code smells in terms of software metrics using feature selection technique based on Correlation on 2896 instances of open-source projects which are extracted from GitHub. Further, we have examined the performance measures like accuracy, precision, F-measure and execution time etc. with the reduced features data set of Android code smells. This paper also discussed about implementation of correlation-based feature selection algorithm to reduce the features of code smells. Then, the data has been analyzed with 4 machine learning algorithms that are Logistic Regression, Stochastic Gradient Descent (SGD), Simple Logistic and Sequential minimal optimization (SMO). The performance metrics for the above-mentioned machine learning algorithms with and without performing the feature selection have been compared. The computed outcome shows that the best accuracy and lesser execution time for all 3 considered Android code smells have been achieved using Logistic Regression algorithm. After feature selection the accuracy has increased up to 16%, 25% and 4.7% for NLMR, MIM and DTWC code smells respectively. Meanwhile, the other performance measures have also been increased.