Technical Debt Forecasting Based on Deep Learning Techniques

Abstract

Technical debt (TD) is a metaphor commonly used to reflect the consequences of quality compromises that can derive short-term benefits but may result in quality decay of software products in the long run. While a broad variety of methods and tools have been proposed over the years for the identification and quantification of TD during the software development cycle, it is not until recently that researchers have turned their interest towards methods aiming to forecast the future TD evolution of a software project. Predicting the future value of TD could facilitate decision-making tasks regarding software maintenance and assist developers and project managers in taking proactive actions regarding TD repayment. In previous relevant studies, time series analysis and Machine Learning techniques have been employed in order to generate meaningful TD forecasts. While these approaches have been proven capable of producing reliable TD predictions, their predictive performance has been observed to decrease significantly for long-term predictions. To this end, in the present paper we investigate whether the adoption of Deep Learning may lead to more accurate long-term TD prediction. For this purpose, Deep Learning models are constructed, evaluated, and compared based on a dataset of five popular real-world software applications. The results of our analysis indicate that the adoption of Deep Learning results in TD forecasting models with sufficient predictive performance up to 150 steps ahead into the future.