A truly human-centred approach to Machine Learning (ML) must consider how to support people modelling phenomena beyond those receiving the bulk of industry and academic attention, including phenomena relevant only to niche communities and for which large datasets may never exist. While deep feature learning is often viewed as a panacea that obviates the task of feature engineering, it may be insufficient to support users with small datasets, novel data sources, and unusual learning problems. We argue that it is therefore necessary to investigate how to support users who are not ML experts in deriving suitable feature representations for new ML problems. We also report on the results of a preliminary study comparing user-driven and automated feature engineering approaches in a sensor-based gesture recognition task.