Exploratory data analysis is an open-ended iterative process, where the goal is to discover new insights. Much of the work to characterise this exploration stems from qualitative research resulting in rich findings, task taxonomies, and conceptual models. In this work, we propose a machine-learning approach where the structure of an exploratory analysis session is automatically learned. Our method, based on Hidden-Markov Models, automatically builds a storyline of past exploration from log data events, that shows key analysis scenarios and the transitions between analysts' hypotheses and research questions. Compared to a clustering method, this approach yields higher accuracy for detecting transitions between analysis scenarios. We argue for incorporating provenance views in exploratory data analysis systems that show, at minimum, the structure and intermediate results of past exploration. Besides helping the reproducibility of the different analyses and their results, this can encourage analysts to reflect upon and ultimately adapt their exploration strategies.