Requirements Modeling: A Use Case Approach to Machine Learning

Introduction

Machine learning-driven software systems (MLD-SS) may be considered more complex than software systems, such as web applications, enterprise systems, etc. This situation can get worse when considering the vital importance of mission-critical systems in various domains, such as healthcare, finance, and smart production. The use of a variety of information technologies, integration methods, and tools, domain-specific processes can add to the complexity of MLD-SS. On the other hand, it is not unusual to see a considerable number of failure stories about Machine Learning (ML) projects. Industrial applications and research studies report that organizations using ML solutions may be at risk of failure or they can fall short of business and organizational objectives (Uysal, 2022a).

One serious issue that is often neglected is to meet the specific requirements of MLD-SS, such as software engineering (SE) requirements, ML-specific requirements, and domain-specific requirements. SE extends its methods and activities beyond data processing and model development, and it focuses on the software development processes. ML, on the other hand, emphasizes the data processing, model development and deployment. Therefore, aligning, and synchronizing data processing and model development with the SE processes is a serious issue (Amershi et al. 2019). Requirements processes, team management, quality management, risk management, and maintenance are also the core aspects both for SE and ML. However, requirements specification for ML is much more challenging (Rahman et al., 2019). Some studies indicate the dynamic and complex context of the design, development, and maintenance of large-scale MLD-SS. ML contexts differ from the plan-driven or agile software development contexts. Therefore, ML and SE workflow management needs to be integrated with the SE best practices (Lwakatare et al., 2020).

A Microsoft team conducted a research study on the SE challenges for ML and AI projects (Amershi et al. 2019). It is found that data discovery and management is complicated and more difficult than the other SE activities. Model customization and reuse require different skills than that are found in SE teams. Since SE and ML modules can be entangled in various ways, ML processes may also be more challenging than the software processes. Therefore, they suggest that the integration of software development infrastructure with ML development would speed up the ML experimentations. Moreover, automating the ML processes, such as data aggregation, synthesizing, feature extraction, and pipeline are the other major concerns for the teams.

Some studies present the SE challenges as they are related with the issues of requirements engineering (RE) of ML (Uysal, 2022b). The systematic mapping study of Villamizar et al. (2021) reveals that the handling customer expectations and lack of validated RE techniques as the main challenges. Giray’s (2021) extensive systematic literature review (SLR) gives the state of art and SE challenges for ML. He adopts a SE perspective on engineering ML systems. Accordingly, forming coherent teams, requirements process, tailoring, and assessing the SE process, managing, and harmonizing the ML system development with software development are the main issues. Fernández et al (2021) conducts a SLR on SE for AI-based systems and finds safety and dependability as the most studied research topics. Software quality and software testing are dominant whereas the areas such as, software maintenance are neglected. They also identify various SE approaches to AI-based systems that are classified according to the Software Engineering Body of Knowledge (SWEBOK) (2014) knowledge areas (KAs). The data-driven and black box nature of ML may be another great challenge.

Kumeno (2019) conducted the first review to map the SE challenges for ML to the SE KAs defined by the SWEBOK standard. The topics in KAs, such as safety, security, validation, and verification are identified as the major challenge topics over the ones in other KAs. Lwakatare et al (2020) focuses on large-scale ML systems in real-world industrial settings. Adaptability and scalability are found as the most often reported SE challenges while safety and privacy are the least reported quality attributes. Nascimento et al (2020) highlights these five of thirteen categories: test, software quality, data, management, model development, and project management (PM). Design, construction, testing, and configuration management KAs have the highest number of solution proposals.