CS5090 Project Report: Bent Parallel Coordinates

The chief difficulty with the parallel coordinates visualization technique is that large data sets become difficult to read due to cluttering of the data. Many techniques have been proposed and implemented to address this problem. A non-algorithmic solution is to simply display the visualization with a larger physical display area. However, a straightforward implementation of this approach may not yield optimal results. I propose a technique called “bent parallel coordinates” to divide the display area into tiles and display the data in a bent path in order to better utilize the space available on large displays. 1 BACKGROUND AND PREVIOUS WORK Parallel coordinates is a technique for enabling the visualization of high-dimensional data which could not be effectively visualized through the use of standard twoor three-dimensional graphs or tables [5][10]. It accomplishes this by displaying each data field as a single vertical axis and laying out these axes horizontally, drawing each data point as a line that intersects each vertical axis in turn (see fig. 1). Because humans are highly visual learners, this technique improves people's ability to analyze high-dimensional data effectively. Fig. 1. A data set with three variables and four dimensions as rendered with parallel coordinates. One of the primary drawbacks of parallel coordinates is that intersecting lines in high-volume data sets can become so cluttered as to render meaningful analysis of the data impossible. Many techniques have been developed to address this problem such as hierarchical clustering [3], angular brushing [4], and filtering by frequency and density information [1], to name a few. 2 BENT PARALLEL COORDINATES I present an alternative approach, called “bent parallel coordinates,” to improve the usability of parallel coordinates by exploiting the advantages of large, high-resolution display areas. Large displays provide a natural solution to the problem of cluttered data by providing more space on which to draw, so it takes a larger data set before the display becomes cluttered. However, utilizing a large display area for parallel coordinates presents new problems. Notably, the standard method of displaying every axis horizontally adjacent to the others may not utilize available space efficiently. Stretching the vertical display area to extreme proportions to fill the available space may make the data difficult to read (see fig. 2). Fig. 2. A naïve approach to rendering parallel coordinates on a large display may make the data difficult to read or result in poor use of available space. I present an alternative technique which divides the display area into nxm tiles and draws the data in a bent path. For example, the data is displayed beginning with the upper-left tile and begins drawing to the right. When the edge of the display area is reached, the algorithm switches directions and begins drawing the data in the new direction, creating a “corner tile” (see fig. 3). The algorithm repeats this process every time it runs out of space, bending each time, until it has drawn every data axis. Fig. 3. An example data set rendered with bent parallel coordinates. The shaded tile represents a corner tile, where the path of the data bends to follow the contours of the display area.