Network-Based Kernel Calculation
The task of kernel density estimation is more complex for networks than it is for classical Euclidean spaces. Network space is fundamentally one-dimensional, since streets are linear, however at nodes (road junctions) the one-dimensional line ‘splits’ (see Fig. 3a) and the domain has a tree-like structure. Furthermore, there are multiple possible routes between two locations on a network; for example, it is possible to circumnavigate a block both clockwise and counter-clockwise. We must therefore redefine our kernel function to take this new spatial representation into account. We ultimately require a solution in which the risk density decays linearly along an edge, so that the network KDE remains comparable with the planar KDE.