Use of Kohonen Self-Organizing Maps and Behavioral Analytics to identify cross - border smuggling activity
Grant M Brown · 2007
Abstract—Risk assessment of movements entering the UK using Kohonen SOMs and decision trees was found to be 292 % more accurate at discerning smuggling than incumbent systems. Inbound freight is currently assessed using a series of risk flags based upon data provided by ferry companies to Government border agencies and a previous offender watchlist. This data constituted the input for the analysis, but presented complex challenges; including proportionally few seizures to train predictive models and a relatively un-diverse “selected for search” population. These factors prompted use of an unsupervised clustering technique to uncover abnormalities in the data and quantify changes in behavior. A supervised technique (a binary decision tree) was also used to enhance existing profiles. Initial tests indicated movements resulting in seizures had a Euclidean distance from their average centroid 10 times greater than the non-seizure mean and only 5 % of vehicles ever moved significantly from their average centroid. Subsequent blind tests “selected ” the riskiest 20 % of movements- from which 39 % of all seizures and 57 % of high value seizures occurring over the same period were identified. Based upon this research, further work has been commissioned to design a system which uses these methods for targeting other movement types. Index Terms — Kohonen Self-Organizing Maps, decision trees, unsupervised analysis, government applications, SAS.