A framework to visualize temporal behavioral relationships in streaming multivariate data

Big Data analysis for scientific data is extremely challenging due to the following features - high resolution, extreme scale, high acquisition rate, multivariate data format and aggregating in the streaming fashion. Therefore, a visual analysis tool that can process, reduce, manipulate and display extreme-scale data is critical for scientists to make the right decision on-site and adjust their measurement strategies during the experiment. The lack of these tools not only severely reduces the scientific throughput, but also impairs our capability for scientific discoveries. In this paper, we describe StreamVis
^ND
- an interactive framework that provides several linked displays designed to reveal multivariate temporal behavior patterns from various perspectives. All of these displays generalize standard visualization paradigms such as line graphs from time samples to time intervals. As such the integral data type of our application is the time interval which we represent as a vector of time samples. Relationships of time intervals are expressed as similarities, possibly warped over time, of pairs of time vectors. These similarities can be among different variables at the same time interval, or different time intervals of the same variable. The former results in a line graph of streaming variables, while the latter results in a new display we called illustrative transform lines of time intervals over the variables. For both displays since the comparative metric is now pairwise similarity, as opposed to absolute value, we require an optimization algorithm, such as multidimensional scaling to perform mapping into display coordinates. Additional displays include a 2D embedding of temporal snapshots of the variables, as well as a 2D embedding of temporal relationships changes among the variables. We demonstrate our system in an environmental pollution diagnostics setting and have obtained encouraging results.