Road traffic accident clustering with categorical attributes

Abstract

Road accidents are considered the result of a complex interplay between road user(s), vehicle(s), infrastructure and environment. To deal with this complexity and to disentangle - as much as possible - the attribute relationships, we here develop an unsupervised categorical model-based accident clustering. This technique enables us firstly to take the typical categorical data aspect into account. Secondly,instead of employing a heuristic measuring the distance between the accidents - as prevailing cluster techniques would do - it uses a more appropriate density-based similarity to assign the accidents to the different clusters. Finally, using all the available data unsupervisedly, the technique aims at an unbiased discovery of the data's inherent sub-structures. The method is applied to the road accident population observed in a Belgian suburban area (Brabant-Walloon). Our model partitioned this population into 5 clusters. Subsequently, all clusters were profiled, pointing out differences regarding time-dependency, type of road user(s), type of collisions, weather conditions, location,... . Since the determinative variables and the variable interplay clearly varied per clusters, they were studied accordingly. This accident examination at cluster level not only confirmed some existing findings but also generated new insights (or issues to get to the bottom of): the 'weekend accidents' are actually all-week accidents, the safety influence of passengers is subjected to weather conditions and the passenger formula, black zones consist mainly out of two accident types, confirmation of attribute relationships findings (e.g. age-gender) appears to be cluster-dependent, ... . Further research and knowledge discovery techniques can be applied within each of the clusters separately.