A Machine Learning Approach to Enhancing Policing Effectiveness During Critical Incidents Involving Gun Violence

The current project will perform an exploratory analysis of the NIBRS and LEOKA datasets with the goal of identifying factors that have significant correlations with law officer injuries and fatalities in gun-related critical incidents. A primary goal of the project is to develop machine learning algorithms such as decision tree classifiers to funnel all incidents involving gun violence through a set of rules such that the conditions that worsen the impact of gun violence are clearly delineated. The project will also formulate similarity measures between police departments, using AI techniques such as collaborative filtering, with the goal of providing them with a benchmarking tool. Police departments will be able to compare their levels of gun-violence related critical incidents with suitable peer departments and in this manner explore avenues for continuous improvement.

When the ratio Y exceeds 4 officers per 1000 population, the incidence of officer fatalities is greatly diminished. While we cannot claim causality from this simple analysis, as a policy lever, there appears to be a critical ratio for department size that is effective in terms of diminishing officer fatalities. While the probability of a FK event is lowest when female officer ratios are less than 5%, this may simply be due to the fact that most police departments fall under this bucket anyway. The best combination, in terms of resulting in the lowest probability of a FK event, is high female officer ratios > 0.15, in combination with large department size (officer rate per 1000 population > 0.004). We define a benchmarking process as below:

  1. First, define a “feature vector” for a police department that profiles the department
  2. Many reasonable definitions for feature vectors are possible, including the number of officers, officer rate per 1000 population and female officer ratio
  3. Find the K-nearest neighbors of the benchmarking department based on Ranking or other similarity metrics. For instance, there is a rich literature on “recommender systems” that are based upon techniques such as collaborative filtering (Netflix recommends movies and Amazon recommends purchases based upon some of these similarity measures)

Results: The analysis of the NIBRS and LEOKA data sets indicates that there is significant potential for improving policing operations by using data analytics to probe gun-related criminal incidents. We conclude with some recommendations and directions for future research:

  1. Ensure consistent and comprehensive data collection across states. Not all states report to NIBRS currently.
  2. Engage police departments and convey the value of having accurate NIBRS/LEOKA information.
  3. Lots of analytical avenues are still available to pursue: e.g., the impact of officer demographics, training budgets and years on the job on officer fatalities.
  4. Apply advanced machine learning techniques such as decision trees to further drill down into the insights obtained from open-source data sets.
  5. Geo-spatial analysis of incidents and studying variations across states

Project Status:
Completed Report

Principal Investigator (PI):
Ram Gopalan, Clinical Associate Professor, School of Business, Rutgers—Camden