Comparing accuracy of geographic profiling by differences in distance measures

Abstract

 The accuracy of geographic profiling for predicting a serial offender's home/base location was compared by using three different distance measures—the Euclidean distance, the Manhattan distance, and the Shortest route distance—using the data of 1,856 crimes committed by 124 residential burglars in Northern Tohoku area of Japan from 2004 to 2015. Logarithmic and the negative exponential coefficients were estimated as the distance decay function for each distance measure by using leave-one-out cross-validation. Also, search areas were calculated to compare the accuracy of geographic profiling. Results of the Friedman's test indicated significant differences in search areas of the three distance measures for the wide area group which consisted of offenders having a long distance between crime locations. The search area when utilizing the Shortest route distance was the smallest for the logarithmic function, whereas the search areas using the Euclidean distance and the Shortest route distance were smaller than the Manhattan distance for the negative exponential function. Results of the narrow area group did not indicate significant differences in search areas for the three distance measures. Therefore, it was concluded that geographic profiling might be improved by using the Shortest route distance when calculating the probability distribution for offenders committing crimes in a wide area that includes many edges, such as rivers, railroads, and mountains, as well as paths such as bridges and railroad crossings.