Optimized Deeplearning Algorithm for Software Defects Prediction

Abstract

Accurate software defect prediction (SDP) helps to enhance the quality of the software by identifying potential flaws early in the development process. However, existing approaches face challenges in achieving reliable predictions. To address this, a novel approach is proposed that combines a two-tier-deep learning framework. The proposed work includes four major phases:(a) pre-processing, (b) Dimensionality reduction, (c) Feature Extraction and (d) Two-fold deep learning-based SDP. The collected raw data is initially pre-processed using a data cleaning approach (handling null values and missing data) and a Decimal scaling normalisation approach. The dimensions of the pre-processed data are reduced using the newly developed Incremental Covariance Principal Component Analysis (ICPCA), and this approach aids in solving the “curse of dimensionality” issue. Then, onto the dimensionally reduced data, the feature extraction is performed using statistical features (standard deviation, skewness, variance, and kurtosis), Mutual information (MI), and Conditional entropy (CE). From the extracted features, the relevant ones are selected using the new Euclidean Distance with Mean Absolute Deviation (ED-MAD). Finally, the SDP (decision making) is carried out using the optimized Two-Fold Deep Learning Framework (O-TFDLF), which encapsulates the RBFN and optimized MLP, respectively. The weight of MLP is fine-tuned using the new Levy Flight Cat Mouse Optimisation (LCMO) method to improve the model's prediction accuracy. The final detected outcome (forecasting the presence/ absence of defect) is acquired from optimized MLP. The implementation has been performed using the MATLAB software. By using certain performance metrics such as Sensitivity, Accuracy, Precision, Specificity and MSE the proposed model’s performance is compared to that of existing models. The accuracy achieved for the proposed model is 93.37%.