Uncertainty reducing and handling strategies in ML development projects

Although prior literature suggested that machine learning (ML) development can suffer strongly from uncertainty, it neglected to unveil the specific uncertainties arising in ML development projects and to understand their impact on the development process. To address this gap, we conduct an exploratory case study based on 62 interviews with ML experts from a multinational software provider. Our study reveals that uncertainty management strategies either target an uncertainty’s reducible or irreducible part and can thus be divided into reducing and handling strategies. We develop a model that shows at which stage of the ML development process each uncertainty is addressed and how as well as by whom the respective reducing and handling strategies are executed. The study mainly contributes to literature on ML development and on uncertainties in software development by unveiling the impact of uncertainty reducing and handling strategies triggered by ML specific uncertainties on the ML development project