Persist: Persistent and Reusable Interactions in Computational Notebooks

Computational notebooks, such as Jupyter, support rich data visualization. However, even when visualizations in notebooks are interactive, they still are a dead end: Interactive data manipulations, such as selections, applying labels, filters, categorizations, or fixes to column or cell values, could be efficiently apply in interactive visual components, but interactive components typically cannot manipulate Python data structures. Furthermore, actions performed in interactive plots are volatile, i.e., they are lost as soon as the cell is re-run, prohibiting reusability and reproducibility. To remedy this, we introduce Persist, a family of techniques to capture and apply interaction provenance to enable persistence of interactions. When interactions manipulate data, we make the transformed data available in dataframes that can be accessed in downstream code cells. We implement our approach as a JupyterLab extension that supports tracking interactions in Vega-Altair plots and in a data table view. Persist can re-execute the interaction provenance when a notebook or a cell is re-executed enabling reproducibility and re-use. We evaluated Persist in a user study targeting data manipulations with 11 participants skilled in Python and Pandas, comparing it to traditional code-based approaches. Participants were consistently faster with Persist, were able to correctly complete more tasks, and expressed a strong preference for Persist.