Crime prediction using spatio-temporal data

Abstract

A crime is an action which constitutes a punishable offence by law. It is harmful for society so as to prevent the criminal activity, it is important to understand crime. Data driven researches are useful to prevent and solve crime. Recent research shows that 50% of the crimes are committed by only handful of offenders. The law enforcement officers need early information about the criminal activity to response and solve the spatio-temporal criminal activity. In this research, supervised learning algorithms are used to predict criminal activity. The proposed data driven system predicts crimes by analyzing San Francisco city criminal activity data set for 12 years. Decision tree and k-nearest neighbor (KNN) algorithms are applied to predict crime. But these two algorithms provide low accuracy in prediction. Then, random forest is applied as an ensemble methods and adaboost is used as a boosting method to increase the accuracy of prediction. However, log-loss is used to measure the performance of classifiers by penalizing false classifications. As the dataset contains highly class imbalance problems, a random undersampling method for random forest algorithm gives the best accuracy. The final accuracy is 99.16% with 0.17% log loss.