HypoVis

Time-oriented data is ubiquitous in many real-world problems from areas like business, medicine, or ecology. Therefore, we consider it very important to propose novel methods for the analysis of time-oriented data. A relatively new approach to solve problems of data analysis is Visual Analytics. The analytical reasoning abilities of automated systems are combined with human reasoning and expert knowledge by the means of interactive visualization. The ultimate goal is gaining insight into the available data, and several important steps in between have been worked out that describe how this goal can be reached. Currently, the discipline mainly consists of two strands: The first one provides insights based on scientific models generated using automated analysis. The second one provides insights based on hypotheses generated using interactive visualizations. The strands are hardly connected in many current frameworks. Combining and intertwining these two strands in an interactive manner would improve the analytic power of the methods available to date. Our goal is the development of a Visual Analytics framework that supports the formulation of scientific models in a comprehensive Visual Analytics process. These models are then to be used for data assessment and forecasting. The most important part of our work is the model definition process using interactive visualizations. To reach this goal, we propose a data model capable of containing such information that can be used by several steps of the Visual Analytics process. Consequently, we go from several distinct process strands to an interactive and iterative process inside an integrated framework. We focus our definition of the data model on the important case of time-oriented data. Problems involving time- oriented data need special attention because time data is different from other kinds of data. Time has an inherent structure, most prominently seen in the calendar aspect of time being composed of smaller granularities, like years and seasons. As these granularities influence natural and social aspects found again in the data, explicitly harnessing the structures of time in data analysis methods can greatly improve the amount of information gained. Therefore, our goal is to include the structure of time in the data model to ease its use in our Visual Analytics framework.

The project HypoVis was conducted for finding methods to formulate hypotheses and validate them with Visual Analytics methods. There can be rather different kinds of hypotheses. For example, they can have a local scope (e.g., Very high traffic in the afternoon can lead either to a continuous decline, or go up again to a short spike in the late evening. Till midnight, it will decline strongly in both cases.). Hypotheses can also have a global scope (e.g., The data from this time series can be approximated using the following statistical formula: ). Hypotheses often start out as an idea in the analysts mind. To validate such hypotheses with a computer, it is necessary to model them in a data structure. After assessing the requirements, we developed a research prototype that is capable of doing this, and much more: TimeBench (examples, documentation, and code download is available at http://www.timebench.org). The main results of HypoVis are Visual Analytics prototypes using TimeBench, which deal with those different kinds of hypotheses/models: (a) In a re-implementation of the GROOVE visualization developed in one of our earlier projects, we improved the color calculation and increased the flexibility of calendar-based calculations and pixel placement. Moreover, the code is highly modular and has already been reused for numerous other Visual Analytics solutions. (b) The models found by an I-Apriori pattern finding algorithm developed over the course of the project, can be shown using various visualizations. One of them is an extended version of Arc Diagrams. The main novelty introduced with HypoVis is that events can have various lengths that are adapted automatically based on the data, and these lengths are also reflected in the visualization. (c) Our novel data model can deal with uncertainties, like indeterminate intervals that have an earliest and latest beginning, an earliest and latest end, and a minimum and maximum duration. TimeBench is capable of mapping this data structure directly to a PlanningLines Visualization. Moreover, TimeBench is continuously expanded with connection protocols to important state-of-the-art software. It already contains a connection to the R statistics environment which is used in an interactive visual interface for user-driven selection of SARIMA models. Our disseminations of this work include: 7 journal papers, 1 conference papers, 7 workshop papers, 2 posters, 1 book chapter, 3 invited papers, 5 master theses, and 3 bachelor thesis.