Estimating Time-Variant Variables' Effects on a Micro Scale for Time-Based Crime Prediction

Crime prediction research has been biased toward spatial factors, and insufficient research has detailed predictions regarding temporal factors. To build an industrially practical crime prediction system, it is necessary to develop a method that considers both temporal and spatial risk factors. This study estimates the effect of time-variant variables such as the day of the week, season, weather, and events for a fundamental analysis of time-based crime prediction. However, because crime data are often imbalanced data with extremely few positive classes, subdividing the data spatially and temporally makes it difficult to estimate the parameters. Therefore, we attempted to solve this problem by dividing the city into clusters with high spatial homogeneity and analyzing each cluster. We observed different effects of time-varying factors for different types of crime.