On the Long-Term Predictive Capability of Data-Driven Software Reliability Model: An Empirical Evaluation

Abstract

In recent years, data-driven software reliability models have been proposed to solve the problematic issues of existing software reliability growth models (i.e., Unrealistic underlying assumptions and model selection problems). However, the previous data-driven approaches mostly focused on sample fitting or next-step prediction without adequate evaluation on their long-term predictive capability. This paper investigates three multi-step-ahead prediction strategies for data-driven software reliability models and compares their predictive performance on failure count data and time between failure data. Then, the model with the outstanding strategy on each data type is compared with conventional software reliability growth models. We found that the Recursive strategy gives better prediction for fault count data, while no strategy is superior to the others for time between failure data. Such data-driven approach with the best input domain showed performance as good as the best one among the software reliability growth models in long-term prediction. These results indicate the applicability of data-driven methods even in long-term prediction and help reliability practitioners to identify an appropriate multi-step prediction strategy for software reliability.