An extension of iStar for Machine Learning requirements by following the PRISE methodology

The rise of Artificial Intelligence (AI) and Deep Learning has led to Machine Learning (ML) becoming a common practice in academia and enterprise. However, a successful ML project requires deep domain knowledge as well as expertise in a plethora of algorithms and data processing techniques. This leads to a stronger dependency and need for communication between developers and stakeholders where numerous requirements come into play. More specifically, in addition to functional requirements such as the output of the model (e.g. classification, clustering or regression), ML projects need to pay special attention to a number of non-functional and quality aspects particular to ML. These include explainability, noise robustness or equity among others. Failure to identify and consider these aspects will lead to inadequate algorithm selection and the failure of the project. In this sense, capturing ML requirements becomes critical. Unfortunately, there is currently an absence of ML requirements modeling approaches. Therefore, in this paper we present the first i* extension for capturing ML requirements and apply it to two real-world projects. Our study covers two main objectives for ML requirements: (i) allows domain experts to specify objectives and quality aspects to be met by the ML solution, and (ii) facilitates the selection and justification of the most adequate ML approaches. Our case studies show that our work enables better ML algorithm selection, preprocessing implementation tailored to each algorithm, and aids in identifying missing data. In addition, they also demonstrate the flexibility of our study to adapt to different domains.

Automated summary

The rise of AI and Deep Learning has made ML a common practice, but successful ML projects require domain knowledge and expertise in algorithms and data processing. This paper presents an approach to capture ML requirements and applies it to real-world projects, showing improved algorithm selection and data preprocessing, and adaptability to different domains.