Quality modeling of machine learning systems
Fraunhofer IESE and Fujitsu Laboratories Limited jointly developed a quality model for quantifying critical properties of machine learning systems in an industrial setting.

Last modified:
Fraunhofer IESE and Fujitsu Laboratories Limited jointly developed a quality model for quantifying critical properties of machine learning systems in an industrial setting.
Last modified:
Artificial Intelligence (AI) is already present in many systems today: from voice assistants to intelligent algorithms that evaluate our online shopping or social media behavior. In the future, we will encounter AI systems much more frequently, especially in critical application areas such as autonomous driving, production automation/Industrie 4.0, or medical technology. The construction and operation of reliable AI systems poses a challenge in practice because many classical quality assurance procedures cannot be applied to AI components or can only be applied to a limited extent: In Machine Learning (ML), for example, functionality is created by applying algorithms to data and is not programmed in the traditional way; the resulting model is hard to understand for humans.
Fujitsu Laboratories Limited is one of the leading companies when it comes to research into explainable AI, meaning that the reasons for decisions are easy for people to understand. In this project, Fraunhofer IESE and Fujitsu Laboratories Limited jointly developed a quality model for ML-based components in a software system. We identified critical quality requirements from different phases in a data science lifecycle and defined relevant quality properties of different entities in an ML-based system (such as the robustness of the trained model, the balancedness of the data used, or the suitability of the infrastructure). We also proposed example measures for quantifying critical properties and proposed a process for the systematic evaluation of measurement data.
Such a quality model helps to objectively evaluate adherence to quality requirements. Furthermore, the process we developed for constructing and tailoring a quality model can be used in different application scenarios.
More details can be found in the following publications:
Julien Siebert, Lisa Joeckel, Jens Heidrich, Adam Trendowicz, Koji Nakamichi, Kyoko Ohashi, Isao Namba, Rieko Yamamoto, Mikio Aoyama. Construction of a Quality Model for Machine Learning Systems. Software Quality Journal. Software Quality Journal 3/2021.
Julien Siebert, Lisa Jöckel, Jens Heidrich, Koji Nakamichi, Kyoko Ohashi, Isao Namba, Rieko Yamamoto, Mikio Aoyama. Towards Guidelines for Assessing Qualities of Machine Learning Systems. QUATIC 2020: 17-31.
Koji Nakamichi, Kyoko Ohashi, Isao Namba, Rieko Yamamoto, Mikio Aoyama, Lisa Jöckel, Julien Siebert, Jens Heidrich. Requirements-Driven Method to Determine Quality Characteristics and Measurements for Machine Learning Software and Its Evaluation. RE 2020: 260-270.