Large-Scale Deep Point Process Models for Crime Forecasting

Statement of the Problem. Recent years have seen a surge of large, complete crime records and related information collected by law enforcement. The volume of data combined with the development of quantitative techniques has boosted research in predictive policing, which helps to prevent crime and evaluate police intervention. To improve the effectiveness of this promising method, a natural question to ask is what makes a good predictive policing model. Given the size of the data, scalable methods are essential for real-time forecasting and evaluation. Moreover, crime records contain rich information aside from spatio-temporal stamps, such as gang involvement, brief description of crime, and intervention attempts. As a result, a multivariate representation of the crime events is useful in modeling and can utilize additional data. Finally, in crime forecasting, the ability to generalize the algorithm is ideal since we are more interested in crimes that we have not seen yet. We propose a unified model that keeps all three properties of a good predictive policing model, including scalability, generalization ability as well as the use of the multivariate model.
Partnerships. We are collaborating with the Gang Reduction & Youth Development program (GRYD) in Los Angeles and Los Angeles Police Department (LAPD), using their experiences in real-world problems and local crime and gang intervention data on our models.
Research Design and Methods. We will extend current self-exciting point process models to nonparametric and multivariate case to incorporate covariates such as gang information and crime description. The estimation of this model will be fast for extreme-scale data. The graph structure of multivariate model allows us to combine graph-based deep neural network to improve the generalization ability.
Analysis. We apply our models on publicly available crime data from major cities and data from LAPD and GRYD that has been scrubbed of all personal identifiers. These data can help us to train our models and test our hypotheses on the dynamics of crime. After that we will validate the generalization ability of our model on new crime data sets. Finally, we want to provide evaluation reports to GRYD gang intervention prgram according to our analysis.
Products, Reports, and Data Archiving. Findings will be published in a thesis, peer-reviewed articles, and presentations at professional conferences. Open-source code of this model will be made available for real-time crime forecasting and the synthetic or even anonymized data will be provided for reproducibility. "Note: This project contains a research and/or development component, as defined in applicable law," and complies with Part 200 Uniform Requirements - 2 CFR 200.210(a)(14).
ca/ncf