Modeling Public Crime Type Using Multinomial Logistic Regression and K-Nearest Neighbor: Pre-and During-Pandemic COVID-19

Abstract

There are common factors that characterize individuals who commit crime. Statistically, relationship between factors could be measured and formed using regression models. While modeling crime rates was widely approached using the ordinary regression model. However, this model is not much capable for categorical response variables such as crime type. In this paper, we are interested to classify crime type using some independent variables using multinomial logistic regression and $K$-Nearest Neighbor ($K$-NN) models. While both are powerful models for classification purposes. The secondary crime data was collected from 2019 (pre-pandemic) and 2020 (during a pandemic) in the police office of Payakumbuh Region, West Sumatra, Indonesia. The results indicate that the crime type was influenced by employment status (unemployed persons) and time occurring (daytime) for both years periods. In the testing data phase, the average of accuracy levels of ​​multinomial logistic regression and $K$-NN are 66.81% and 73.86%, respectively. In this case study, $K$-NN model is better approach to be used for the prediction and classification of crime type if compared with multinomial logistic regression. Both models could be considered for supporting the police divisions on decision making and prevention strategy.