Repo for ML Final Project CAPP 30254 Mariel Wiechers, Domingo Carbone, Gabe Morrison
Spring 2021
For this project, we sought to explore predictive policing and its challenges from both machine learning and public policy perspectives. While widely used, many predictive policing algorithms are flawed and racist. Through this project, we hoped to explore these dynamics and specifically consider the extent to which some algorithms could suffer from model drift as a consequence of the 2020 pandemic changing crime dynamics.
This README is organized as follows. We first share the Executive Summary of our final research project and link to our report. We subsequently describe repo organization and general mechanics of the code.
Predictive policing with machine learning algorithms is widely used by police departments throughout the U.S., in an attempt to determine deployment of police resources. Because of the popularity of such approaches, it’s relevant to understand their limitations and possible negative consequences, particularly with respect to handling model drift. This paper builds models to predict the police beats with highest crime incidence in the city of Chicago for the year 2020, based on previous years’ data. It then measures model drift by comparing the chosen models’ 2020 precision and recall to those figures in earlier years. When optimizing for precision, the models remain extremely precise, but there are substantial declines in recall, suggesting model drift occurs in the 2020 data.
You can access our final report here.
Alternatively, it is here: https://docs.google.com/document/d/1zT_j5JFn-bzO95at9PQfqy0UbNwk49TTTVi-SU7H_qI/edit?usp=sharing.
We organize data for the project into raw_data, intermediate_data, and final_data folders. raw_data contains unprocessed data that was downloaded and analyzed. intermediate_data contain processed data upon which we performed machine learning. final_data contains results from models.
Our scripts folder contains the ipynbs we wrote to process and analyze our data and visualize our results. Most files are the names of the ML algorithms they run, but we write helper functions for data cleaning in geoprocessing and geoprocessing_helpers. We visualize crime trends in time_frame_cuts. We also perform data manipulations in a number of scripts including spatial_lag, business_data_cleaning, and data_prep, and visualize results in data_viz.Rmd, graphs_1, and graphs_2.