Crime Prediction in Edmonton Neighbourhoods
Supervised learning is widely used in crime prediction. Both criminology and sociology are sophisticated subjects and require domain knowledge and experience in order to estimate what types of crime will occur in a neighborhood. Because of the correlation among different categories of crime, we apply multitask learning to predict crime. In this project, we explore three different machine learning algorithms and measure their performance on predicting the number of crimes using root mean squared error. The datasets are from the City of Edmonton’s Crime and Census data from 2012 to 2016. We apply multitask least square support vector regression, the Curds and Whey algorithm and multitask random forest to our dataset. Our approach outperforms conventional single task learning models by 5 to 8 percent.