Visard: System Architecture

This project centers on discovering the changing patterns over time in linked data and providing a visualization tool for users, which help users to establish historical context and make predictions about upcoming actions.

 Figure 1 shows the architecture of the system. At the lowest layer is the "raw" relational data (which may in fact have been processed to extract metadata and semantic meaning). Model induction methods are used to construct the next layer, a probabilistic relational model that summarizes the correlations and dependencies in the data. This model feeds into a spatial layout layer that uses interactive, incremental hierarchical clustering methods to create a display space. Finally, the dynamic visualization layer allows the user to interactively explore the model in the layout space.

Note that the user has both spatial and temporal control over the visualization. That is, the user can select particular objects and relations to visualize and establish their initial spatial arrangement (with the assistance of interactive, semi-automated layout tools that we will develop). Given the objects and relations that are of interest to the user, probabilistic models will be learned for each available time interval of data. Changes in the data will be reflected in changes in the model, which will in turn be reflected in the visualization. The visualization shows different types of objects and links (known vs. postulated, probabilistic vs. deterministic, temporal vs. static) using visual features such as color and size. Users can control the appearance of the display, allowing them to construct different views depending on heir current task requirements. The user can also control the rate and direction of change (i.e., forward or backward in time), can engage in "what-if" analysis to see the possible long-term effects of hypothetical changes to the data and to compare the time course of the data with and without the change, and can drill down to examine the raw data that supports the model.

We propose  to represent change in the visualization by motion because of its perceptual and cognitive importance. Motion is a so-called "pop-out" visual feature, one that registers pre-attentively. This makes it possible for the user to consciously attend to a particular aspect of the visualization while remaining informed about changes in other aspects. The phenomenon of spatial memory is exploited by the interactive layout clustering methods to guide the construction of the display space, allowing users to focus on those aspects of the linked structure they deem to be most relevant. For example, a user can easily remember that a particular category of individuals (e.g., high-tech startups) is located in the upper right of the display, notice increased activity in that part of the screen out of the corner of their eye, and immediately conclude that the time period covered by the visualization entails significant change within that category.