Software defect prediction model based on improved LLE-SVM

A recent study namely software defect prediction model based on Local Linear Embedding and Support Vector Machines (LLE-SVM) has indicated that Support Vector Regression (SVR) has an interesting potential in the field of software defect prediction. However, the parameters optimization of LLE-SVM model is computationally expensive by using the grid search algorithm, resulting in a lower efficiency of the model; and it ignores the imbalance of data sets when using SVM classier to differentiate the defective class and non-defective class. Thus resulting in a lower prediction accuracy. To solve these problems in LLE-SVM model, we propose a new software defect prediction model based on the improved Locally Linear Embedding and Support Vector Machines (ILLE-SVM). ILLE-SVM model employed the coarse-To-fine grid search algorithm to search the optimal parameters. It ensured a high accuracy of the parameters and reduced the parameters optimizing time by gradually narrowing the search scope and enlarging the parameters step. As for the question that SVM suffers a performance bias in classification when data sets are unbalanced, we employed gird search algorithm to automatically set the reasonable weights of different class. The comparison between LLE-SVM model and ILLE-SVM model is experimentally verified on four NASA defect data sets. The results indicate that ILLE-SVM model can search the optimal parameters faster than LLE-SVM model and perform better than LLE-SVM in software defect prediction.