Chicago Crime Scene
Arrest Prediction
Chicago Crime Scene
Arrest Prediction
Chicago Crime Scene
Project Goals
- Goal: Provide predictive guidance of likelihood of arrest.
- Plan: determine rate of arrest per crime reports.
- Method: use the dataset from the Chicago Data Portal.
Chicago Data Portal
- Dataset available via the Socrata Open Data API
(and bulk downloads). - Developed a SODA to SQL schema converter for import.
- SODA access permits daily updates.
Initial Models
- Initial modeling seemed to show performance above baseline (of 80% accuracy).
- Client wasn't necessarily impressed.
Much about Metrics
- Accuracy: Client already knows "most of the time" they don't get arrested.
- Concern #1: If I say no arrest, and I get that wrong, they lose "staff". I need to be Sensitive of that.
- Concern #2: If I say arrest, and I get that wrong, they lose "business". So I need to be Precise.
Crime by Communities
- The dataset calls these "Location Areas".
- These are the colloquial parts of town.
- This is also the most general breakout.
Crime by Wards
- The political districts
- A bit more granular
- Politicians draw much more convoluted shapes.
Crime by the Beat
- Distrcit and Beat per the CPD.
- District and Beat are they only purely heirarchical geo features.
- Beat is much more granular.
Chicago Crime Scene
- Geo Data was unpersuasive.
- Time Data was meager.
- Two remaining features proved very useful:
- Location DESCRIPTION
- Crime DESCRIPTION
Chicago Crime Scene
Model Results
Arrest Predictor
Appendix
Chicago Crime Scene
Chicago Crime Scene
Chicago Crime Scene
Chicago Crime Scene
Chicago Crime Scene
Chicago Crime Scene
