1. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning Authors: Chung-Hsien Yu 1, Wei Ding 1, Ping Chen 1, and Melissa Morabito 2 1 Department of Computer Science, University of Massachusetts Boston 2 The School of Criminology and Justice Studies, University of Massachusetts Lowell PAKDD 2014 Presented by: Dr. Wei Ding

2. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 A new concept of multi-dimensional feature denoted as spatio- temporal pattern (STP) is proposed. This kind of STP is extracted from crime distribution in different time periods at different granularity levels and treated as features. The Cluster-Confidence-Rate-Boosting (CCRBoost) algorithm is designed to efficiently select relevant STP to construct a global ensemble crime pattern. This ensemble crime pattern is capable of predicting future crime. Our case study shows that CCRBoost algorithm has achieved about 80% on accuracy. Abstract 2

3. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime Data 3 Source: www.crimereports.com

4. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime STP - Distribution 4

5. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime STP Construction 5

6. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 STP - Non-stationary 6

7. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Which of these patterns should be selected to form the global crime pattern? How can this global ensemble pattern be constructed? Challenges 7

8. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 CCRBoost 8

9. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime Pattern - Seasonal 9 Source: Cohen, J., Gorr, W.L.: Development of crime forecasting and mapping systems for use by police. H. John Heinz III School of Public Policy and Management. Carnegie Mellon University (2005)

10. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime Pattern - Density 10 Source: Short, M.B., Bertozzi, A.L., Brantingham, P.J.: Nonlinear patterns in urban crime: Hotspots, bifurcations, and suppression. SIAM Journal on Applied Dynamical Systems 9(2), 462–483 (2010)

11. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Crime Pattern - Boundaries 11 Source: Kumar, M.V., Chandrasekar, C.: Spatial clustering simulation on analysis of spatialtemporal crime hotspot for predicting crime activities. International Journal of Computer Science and Information Technologies 2(6), 2864–2867 (2011)

12. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Training Error 12

13. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Confidence Value 13 Our Loss Function

14. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Objective Function 14 We want to find the minimum value of Z to lower the training error.

15. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Minimum Z Value 15 first derivative 2nd derivative

16. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Update Weights 16 The goal is to exponentially lower the weights on those vectors that are recognized by the pattern.

17. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Strong Hypothesis 17 The final global STP is defined as: α is a user-defined threshold

18. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 4-years’ crime records have been used for the evaluation. In addition, three different grid resolutions have been applied. We are targeting the residential burglary crime. Arrest, commercial burglary, foreclosure, motor vehicle larceny, 911 call, and street robbery records are used as the indicators. LADTree is chosen to extracted the local STP. Case Study 18

19. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Comparing with Random Sampling 19

20. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Comparing with Other Methods 20

21. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Convergence 21

22. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 The Resulting Global STP 22

23. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Based on the consistency with actual crime patterns, our algorithm does find the patterns which recognize not only the spatial but also the temporal factors that are useful for criminal justice professionals in predicting the incidence of future crime. From the visualization of the resulting STP, the patterns selected from this algorithm are indicative of the true locations of residential burglaries throughout the target city. The ultimate goal of our research is to build a crime prediction system with strong predictive power, which is able to provide forecast in a timely manner and requires less amount of data inputs. SUMMARY 23

24. Crime Forecasting Using Spatio-temporal Pattern with Ensemble Learning PAKDD 2014 Q & A THANK YOU!! 24 Chung-Hsien Yu (csyu@cs.umb.edu)csyu@cs.umb.edu Wei Ding(ding@cs.umb.edu)ding@cs.umb.edu