Performance Evaluation of Support Vector Classifier Kernel Functions in Method-Level Refactoring Analysis

Refactoring Techniques work as effective tools for improving code quality without altering its external behavior. The main objective of this work is to analyze the method level of refactoring by using a support vector classifier and kernel function. We have implemented different kernel functions in three open-source Java projects (antlr4, oryx & titan). The study aims to understand how the support vector algorithm (SVC) works for different kernel functions (linear, poly, RBF & sigmoid) related to code improvements. It also deals with the problem of unbalanced data by using both under-sampling (NearMiss) and oversampling (Random oversampling example (Rose)) techniques to balance the dataset. After that, we focus on how well the SVC of different kernel functions works under each sampling method. Result: Our analysis reveals that the mean accuracy for the NearMiss and Rose data sampling techniques is 100 % for the classifier SVC (linear) and 95% (sigmoid) respectively.