The Materiality of Data: Practices of Coping with Imperfect Data for Machine Learning Development

We study how the materiality of data shapes the practices of data work. Empirically, we look at the development of machine learning models for predicting complex agricultural phenomena. We find that representing phenomena by data involves an iterative process of data workers redefining phenomena to make them representable by data, and reconstructing data modulated by the material aspects of data’s recording. In particular, we find four practices data workers enact as they cope with these challenges of representing complex phenomena: transforming data, resourcing new data, redefining phenomena, and proxy making. The paper contributes to the literature on data work by showing the consequential role of materiality in data work which enables us to develop a taxonomy for systematic analysis of data work practices and introduce a novel practice of data work - proxy making.