Approximating scatterplots of large datasets using distribution splats

Many situations exist where the plotting of large data sets with categorical attributes is desired in a 3D coordinate system. For example, a marketing company may conduct a survey involving one million subjects and then plot peoples favorite car type against their weight, height and annual income. Scatter point plotting, in which each point is individually plotted at its corresponding cartesian location using a defined primitive, is usually used to render a plot of this type. If the dependent variable is continuous, eve can discretize the three-dimensional space into bins or voxels and retain the average value of all records falling within each voxel. Previous work employed volume rendering techniques, in particular, splatting, to represent this aggregated data, by mapping each average value to a representative color. However, if the color mapped attribute is categorical, then this technique is inadequate, since we can not represent or calculate an average value to which a single color is assigned. This paper presents a new technique called distribution splatting. Our method is not only faster than traditional scatter plotting for massive data sets, but also visually preserves the sample density within the volume. The main problem examined in this paper, is the representation of all the attributes at a voxel using a single splat. We represent a splat with a multi-colored tessellated hexagon. The number of elements, opacity and size within each splat can be modulated with respect to record size to show the sample density accurately. Other techniques such as percentage thresholding and an evidence viewing mode are presented as extensions. Examples from real data sets are presented to display the effectiveness of our method.