Impacto da latência na verificação e do atraso acentuado na rotulação para detecção de falhas de software

Abstract

Software Defect Prediction is an activity inherent to software development and it requires a high amount of human effort. This activity is often not prioritized in order toreduce the project's expenses. Just-in-Time Software Defect Prediction (JIT-SDP) is one of the approaches used for predicting software defects in order to automatically identify, through machine learning methods, software artifacts likely to contain defects based on historical data. However, most approaches assume that the characteristics of the problem remain the same over time, but the imbalance between classes is a problem that evolves over time as new training examples arrive, for example. Analyzing the problem in an online environment means that, among other things, there is an intrinsic chronological aspect to be considered which, in turn, brings with it some issues, among them verification latency, which refers to the fact that training example labels can arrive much later than their characteristics. This work aims to investigate the impact of verification latency on the problem of detecting defects in software, as well as the performance of these methods according to the degree of verification latency in examples of the defect inducing class.