Kernel Density Estimation-based Edge Bundling

Edge bundling is a recent, increasingly promising, technique which generates graph layouts of limited clutter. Bundled layouts can be used to get insight into the coarse-scale structure of networks, geographical maps, and software systems.

For general graphs, many bundling methods have been proposed in the last few years. However, the following requirements are still challenging

bundle graphs of tens..hundreds of thousands of edges efficiently (near-real-time)
declutter graphs with many overlapping edges and nodes
intuitively control the look and feel of the bundling (e.g. produce smooth or ramified bundles)
easy implementation (no complex parameter settings or algorithms)

Kernel density estimation

We present here a method that complies well with the above requirements. The principle of our method is simple. Given an initial graph drawing

convert the drawing to a density map using kernel density estimation (KDE)
compute the normalized density map gradient
move each edge in the gradient direction
smooth edges using Laplacian filtering (optionally)
repeat from step 1 with decreasing kernel sizes

Intuitively, the above is equivalent to sharpening the edges' density map. This in turn pulls edges towards the center of their local point spatial distribution, which achieves the bundling.

Implementation

KDEEB is simple to implement and can be easily accelerated using texture splatting for the computation of density maps and their gradients. Our entire implementation is done in C# using OpenGL 1.1.

An implementation of KDEEB will be soon available here.

See also CUBu, our even faster CUDA-based version of KDEEB.

Results

Below are shown several bundling results obtained with KDEEB. The input graphs used are well-known from other Infovis research papers. For comparison purposes, layouts of the same graphs obtained with other recent bundling methods are shown:

FDEB: Force-directed edge bundling
WR: Winding roads edge bundling
MINGLE: Multilevel agglomerative edge bundling
SBEB: Skeleton-based edge bundling
KDEEB: Our method

Note: You can download the images below (click, save as..) to see higher-resolution versions.