Modeling and Assessing User Interaction in Big Data Visualization Systems

When users interact with large data through a visualization system, its response time is crucial in keeping engagement and efficacy as high as possible, and latencies as low as 500 ms can be detrimental to the correct execution of the analysis. This can be due to several causes: (i) for large data or high query rates, database management systems (DBMS) may fail to meet the performance needs; (ii) modeling all the interactions with a visualization system is challenging due to their exploratory nature, where not all of them are equally demanding in terms of computation time; (iii) there is a lack of models for integrating optimizations in a holistic way, hampering consistent evaluation across systems. In response to these problems, we propose a conceptual interaction-driven framework that enhances the visualization pipeline by adding a new Translation layer between the Data-, Visualization- and Interaction- layers, leveraging the modeling of interactions with augmented statecharts. This new layer aims to collect information about queries and rendering computations, linking such values to interactions in the statechart. To make the Translation layer actionable, we contribute a software component to automatically model the user interactions for a generic web-based visualization system through augmented statecharts, in which each interaction is labeled with its latency threshold. We first demonstrate its generality on ten state-of-the-art visualization systems. Then we perform a user study (n = 50), collecting traces by asking users to perform already established exploratory tasks on the well-known Crossfilter interface. Finally, we replay those traces over its generated statechart, assessing the capability to model the user interactions correctly and describing violations in the latency thresholds.