A Study of Applying Deep Learning-Based Weighted Combinations to Improve Defect Prediction Accuracy and Effectiveness

Software errors or bugs are the primary cause of poor software quality. Thus defect prediction is a prominent approach to enhance software quality. It is a common technique for identifying defect-prone programs, which help the practitioners allocate needed quality assurance efforts (e.g., testing and debugging). An accurate prediction may bring significant benefits. However, there is still space for improvements by applying different levels of instances or using some state-of-the-art techniques to construct the prediction models. In this paper, we propose a weighted combination method of activation functions to improve the effectiveness of defect prediction, comprising of weighted arithmetic, geometric, harmonic, contra-harmonic, and cubic combinations. When there are several kinds of classifiers, the method of weighted combinations can be applied to combine the strengths and create a new model or classifier. That is, weighted combination method(s) would be able to combine the advantage of different activation functions to train the neural network in deep learning. Six open-source projects are used to to evaluate the performance of weighted combination methods of activation functions: single, double- and triple-weighted approaches. Experimental results show that double-weighted combinations of activation function outperform single and triple-weighted combinations. It is also worth noting that single activation function outperform triple-weighted combination.