Software defect prediction is a practical approach to improve the quality and efficiency of time and costs for software testing by focusing on defect modules. The dataset of software defect prediction naturally has a class imbalance problem with very few defective modules compared to non-defective modules. This situation has a negative impact on the Neural Network, which can lead to overfitting and poor accuracy. Synthetic Minority Over-sampling Technique (SMOTE) is one of the popular techniques that can solve the problem of class imbalance. However, Neural Network and SMOTE both have hyperparameters which must be determined by the user before the modelling process. In this study, we applied the Neural Networks Based SMOTE, a combination of Neural Network and SMOTE with each hyperparameter of SMOTE and Neural Network that are optimized using random search to solve the class imbalance problem in the six NASA datasets. The results use a 5*5 cross-validation show that increases Bal by 25.48% and Recall by 45.99% compared to the original Neural Network. We also compare the performance of Neural Network-based SMOTE with “Traditional” Machine Learning-based SMOTE. The Neural Network-based SMOTE takes first place in the average rank.