Estimating the risk of insect species invasion: Kohonen self-organising maps versus k-means clustering

Previous work on the estimation of the invasiveness of insect pest species used a single Kohonen self-organising map (SOM) to quantify the invasion potential of each member of a set of species in relation to a particular geographic region. In this paper that method is critically compared to an alternative approach of calculating the invasive potential of insect pest species as an outcome of clustering of regional species assemblages. Data clustering was performed using SOM and k-means optimisation clustering and multiple trials were performed with each algorithm. The outcomes of these two approaches were evaluated and compared to the previously published results obtained from a single SOM. The results show firstly, due to the inherent variation between trials of the algorithms used, that multiple trials are necessary to determine reliable risk ratings, and secondly, that k-means clustering can be considered a more appropriate algorithm for this particular application, as it produces clusters of higher quality, as determined by objective cluster measures, and is far more computationally efficient than SOM.