Developing an Intelligent Model for Technical Debt Management in Large-Scale Software Projects Using Deep Learning and Predictive Analytics
TL;DR
According to the results, deep learning can be more beneficial for forecasting technical debt but does not bring the required level of improvement only due to the complexity of the model.
Abstract
Technical debt (TD) is an everlasting issue in software engineering. This is the issue of short-term programming choices leading to larger costs in terms of maintenance, quality, and evolution in the future. This issue becomes critically important for large software projects, as they usually contain various quality indicators and constantly changing states of the system. The goal of the study is to create and evaluate an efficient intelligent model for forecasting technical debt. The experiment utilized the Technical Debt Forecasting dataset available on Zenodo. The dataset has been gathered from 15 open source programming projects. The compiled dataset is formed by 1,918 measurements taken during the process of the software under consideration. The dataset consists of 41 common predictive variables. The report uses 70/15/15 temporal split to avoid time leakage. The TCN utilizes an eight-step temporal window and outperforms persistence, Random Forest, and Gradient Boost models. The predictions of normalized technical debt for the second step based on TCN in the within-project test observations (n = 172) yielded MAE = 0.001388, RMSE = 0.002956, and R² = 0.999842. The results indicated that on this dataset, the persistence method achieved better scores of MAE = 0.001166, RMSE = 0.002943, and R² = 0.999843. To analyze what the differences in results are in terms of statistical significance, a Wilcoxon signed-rank test was applied to compare paired absolute errors, and the computed differences turned out to be statistically significant (p < .001) favoring the persistence model. The TCN was also used for getting results on cross-project data with Groovy, Kafka, and CommonsIO being excluded. The results there demonstrated MAE = 0.003930 for TCN compared to 0.003110 for persistence. According to the results, deep learning can be more beneficial for forecasting technical debt but does not bring the required level of improvement only due to the complexity of the model.
