Application of Recurrent Neural Networks to Covariate Software Defect Prediction  
Author Fatemeh Salboukh


Co-Author(s) Priscila Silva; Vidhyashree Nagaraju; Lance Fiondella


Abstract Traditional software reliability growth models (SRGM) assess defect discovery as a function of testing time or effort to quantify failure intensity based on the Non-Homogeneous Poisson Process (NHPP). More recently, covariate NHPP models have substantially improved defect prediction by characterizing software defect discovery in terms of underlying testing activities. Traditional NHPP models are based on parametric forms with specific distributional assumptions, which often fail to capture all details present on defect data. Therefore, this paper assesses the effectiveness of neural networks to predict software defects including covariates to encode factors driving defect discovery explicitly. Two neural networks are considered, including recurrent neural networks (RNNs) and long short-term memory (LSTM), which are then compared with the traditional covariate models to validate predictive capability. Our results suggest that LSTM achieves better overall goodness-of-fit measures, such as approximately 6 and 9 times smaller mean square errors and mean absolute percentage error, respectively, compared to traditional models when 75% of the data is used for training. These results suggest that neural networks are able to track and predict defect trends more accurately than traditional methods.


Keywords Software reliability, non-homogeneous Poisson process, covariate models, defect discovery, neural networks
    Article #:  RQD28-10

Proceedings of 28th ISSAT International Conference on Reliability & Quality in Design
August 3-5, 2023