Forecasting software system quality metrics using modifications of the DeepLIFT interpretation method

• Published in: Sučasnij stan naukovih doslìdženʹ ta tehnologìj v promislovostì (Online) no. 4(30); pp. 110 - 124
• Main Authors: Shantyr, Anton, Zinchenko, Olha
• Format: Journal Article
• Language: English
• Published: 11.12.2024
• ISSN: 2522-9818; 2524-2296
• DOI: 10.30837/2522-9818.2024.4.110

The subject of the study is to improve the interpretative method of DeepLIFT (Deep Learning Important Features) in the context of ensuring quality assessment of software systems (PS). The algorithmic and mathematical aspects of the DeepLIFT base method, as well as its improvement for analyzing and improving the quality of PS, are being studied. Purpose: development and justification of five methods of optimization modifications of the DeepLIFT method to improve the accuracy and efficiency of forecasting the quality of PS quality. The article solves the following tasks: to conduct a detailed review of problems related to the use of DeepLIFT when assessing the quality of the aircraft; provide a mathematical description of the five ways of modifying the DeepLIFT method aimed at improving the accuracy, adaptability and speed of quality assessment; experimental testing of the proposed modifications is performed to evaluate their effectiveness in assessing the quality of aircraft compared to the original DeepLIFT method. Research methods: analysis of literature; methods of experimental verification; calculation of average values and use of significance tests; modeling. The results were achieved: 1) a detailed analysis of the limitations of the base method of DeepLIFT in the context of modern software systems, which revealed low adaptability to dynamic data, limited interpretability for complex architectures and difficulty with stability of results in cases of variable operating conditions; 2) five DeepLIFT method is proposed; 3) the proposed modifications have undergone experimental testing, which demonstrated their effectiveness compared to the original DeepLIFT method. The results of the study showed that all modifications allow you to achieve improvements in key parameters. Conclusions: in the face of increasing the complexity of architectures, increasing the volume of data and productivity requirements, innovative methods of evaluation and predicting the quality of PS become a necessary. Each of the improved methods increases accuracy and interpretability compared to the base DeepLIFT. The highest accuracy is shown by the temporal DeepLIFT (T-Dlift) respectively.