ARSRPE Conference Paper Database

Decision tree analysis: Prediction of serious traffic offending

Gosse, Michelle



The objective was to predict whether an offender would commit a traffic offence involving death, using decision tree analysis.  Four groups of predictive models were produced; two based on serious offending only (i.e. traffic offences that escalated to Court) and two based on time-bound (i.e. Period) data constructed from all traffic offending.  The small number of "risk" offenders who committed a traffic offence involving death - 1.44% of the Court-based records and 0.14% of Period-based records - led to the use of a profit matrix in an attempt to mitigate against the small number of target records.  While none of the predictive models were particularly useful in a practical sense, the non-aggregated Court data produced the best result as some individual nodes were more useful than others.  The decision tree models generated for the Period data were least useful, although this could be due to the much lower number of "risk" offenders and the relatively short timeframe covered by the Period data.  A key finding is that aggregated data based on predefined Police traffic offence groups appears unsuitable for decision tree analysis.