Comparison of Self-Organizing Maps, Mixture, K-means and Hybrid Approaches to Risk Classification of Passive Railway Crossings

We create factor constructs for the physical characteristics of 864 passive railway crossings in South Australia. Crossings are then classified as dangerous or otherwise by means of self-organizing maps, K-means, mixture models and combinations of the latter. Results are compared using historical accident data. The self-organizing map with mixtures approach is found to be optimal in prediction of dangerous crossings. Results show that there exists two types of dangerous crossings. This has significant implications for prioritisation of crossing upgrades.