Mosca, AbigailRobinson, ShannonClarke, MeredithRedelmeier, RebeccaCoates, SebastianCashman, DylanChang, RemcoJohansson, Jimmy and Sadlo, Filip and Marai, G. Elisabeta2019-06-022019-06-022019978-3-03868-090-1https://doi.org/10.2312/evs.20191173https://diglib.eg.org:443/handle/10.2312/evs20191173As the sophistication of data analyses increases many subject matter experts looking to make data-driven decisions turn to data scientists to help with their data analysis needs. These subject matter experts may have little to no experience in data analysis, and may have little to no idea of what exactly they need to support their decision making. It is up to data scientists to determine the exact analysis needs of these clients before they can run an analysis. We call this step of the analysis process initialization and define it as: translating clients' broad, high-level questions into analytic queries. Despite the fact that this can be a very time consuming task for data scientists, few visualization tools exist to support it. To provide guidance on how future tools may fill this gap, we conducted 14 semi-structured interviews with client-facing data scientists in an array of fields. In analyzing interviews we find data scientists generally employ three methods for initialization: working backwards, probing, and recommending. We discus existing techniques that share synergy with each of these methods and could be leveraged in the design of future visualization tools to support initialization.Humancentered computingHuman computer interaction (HCI)Visualization10.2312/evs.2019117373-77