Most users are able to obtain their explorable ideas for a data table but cannot clearly declare their analysis tasks as visual queries. Visualization recommendation methods can reduce the demand for data and design knowledge of visual analysis by extracting or referring information in existing high-quality views.

However, most solutions need users to articulate a precise constraint when they try to steer recommendations with their intentions and insights. To address this limitation, we allowed participants to create and alter multiple views for data exploration in a Workshop to further examine the analytical requirements entailed by users constantly generating charts. Our deep learning model was built using triangulation from Workshop to dynamically perceive the users’ analytical tasks in the editing sequence on data and views.

We provided Dowsing, a mixed-initiative recommendation system for multiple-view visualizations. It can utilize and expose the user’s potential analysis tasks to recommend visualizations during the explorative building phase. Meanwhile, Dowsing allows users to confirm and edit their intentions as they explore further to quickly adapt to changing the analysis requirements. We evaluated our deep learning model through quantitative experiments and verified the effectiveness of Dowsing by user study using vega-datasets.

The workflow of Dowsing is shown as above. The system can be devided into two main parts: the perception and the recommendation generation. The preception uses LSTM model, together with statistical methods, to predict user intent. The user intent, represented as a set of constraints, is mixed with data facts and generic visualization rules and used together with the ASP solver in recommendation generation, which in turn yields a series of recommended views.

You can watch our overview video below.