Software Defect Prediction using Multi-Layer Perceptron based Support Vector Machine Classifier

In recent days, Software Defect Prediction (SDP) systems are evolving gradually as they improve software reliability and quality by ascertaining the defects in software. However, the testers identify the buggy codes, but do not detect the hidden bugs as it delays the production. In existing systems, the SVM classifier is used in several existing models, which has raised multiple problems such as robustness, slow learning rate and accuracy. In this paper, a Machine Learning (ML) approach, named as Multi-Layer Perceptron based Support Vector Machine (MLP-SVM) is proposed for SDP. The SVM classifier achieved robustness with the help of hinge loss and increased the learning rate with SGD weight optimization. Initially, the input is taken from the PROMISE repository and preprocessed to create tokens with the help of tokenization for eliminating comments and blank spaces. After that, Bidirectional Encoder Representation Transformer (BERT) is employed for the input embedding to convert sequence of tokens into vector-indexed sequences. Then, these index sequences are processed into SVM regression activation layer is used for classification to classify defect and non-defect codes. From the results, the proposed MLP-SVM achieved better results in terms of accuracy, precision, recall, f1-score and AUC are 89.23%, 72.50%, 75.33%, 87.57% and 73.29% respectively than existing transfer learning model.