Improving Public Safety with Machine Learning

Machine learning usage to increase public safety by increasing the response of the NYPD to gunfire by solving the main restrictions the government operated

The biggest mission of NYPD if to increase public safety by effectively allocating public, limited, resources. In the criminal landscape, firearms incidents play a significant role in public safety and is one of the main focuses of the police department. In New York City, firearms are used in 68% of murders and 41% of robberies1. With a limited visibility of the all incidents that occurs in NYC, the government relies in the proactivity of the community to respond to gunshot incidents that may lead to a low efficiency in human resource allocation and effectiveness of further investigations

Increasing efficiency by increasing response to incidents

Since the end of 2016, the NYPD began the roll out of a system called ShotSpotter (SST), a technology that process sounds from acoustic sensors placed in the streets to identify and locate gunfire incidents in real time1. Moreover, SST matched the series of sounds to its existing database of gunfire and classify the sound as a gunfire incident or not, after a positive identification, the SST traces the original location and send the information to the police department.

This technology aimed to increase the response of the department to gunfire by solving the main restrictions the government operated. Until ShotSpotter, the department relies on “911 calls” to act and therefore has limited visibility of the occurrences of gun related incidents in the city.

  • NYPD reports state that approximately 80% of the gunshots are not reported1: The community often are discouraged reporting the incident due to fear of the investigation or the belief that they could handle the situation themselves.
  • In a dense city as NY, Shotgun sounds to identify and determinate its origin: An interview with the CEO of ShotSpotter shows the challenge of correctly identifying the specific sound pattern of a gunshot2 (see link below). In addition, imprecise 911 calls lead to a larger response time to the incident.

Initial Results:  By 2017, only 66% of the incidents reported by ShotSpotter weren’t reported through “911 calls” and the response time for ShotSpotter identification was only 8 minutes lower than any other response time in the department 3,4.

New way of investigating incidents

The implementation of SST, changed the way NYPD structured their operation and introduced a new integrated way of investigating gunshot incidents. The integration of the SST sound database and the NYPD internal crime database, Domain Awareness System (DAS), enabled the department to connect incidents with ongoing investigations (see picture below) even when the incident doesn`t have a physical evidence (e.g. bullet, victim, gun, etc)

Image:  “The way moving forward” NYPD Publication (2016)

Mastering predictive analysis to prevent crime

The amount of data storage that SST and DAS, opens a wide range of possibilities to improve services via machine learning. Recently, police departments across US are working on developing predictive models that goes beyond the reactive way of combating crime. The innovating efforts have a new focus of using the ability of foreseeing incidents as a way of preventing crime.

In the next years, predictive analysis (mass) implementation is around identifying and mapping criminality profile of different regions. In the past year, the NYPD has been more actively introducing tools such as Hunchlab, or ShotSpotter Mission, uses historical data to maps each shift regions based on its risk assessment (types of events such as robbery, homicide, etc) and probability of having an incident in that period. The mapping based of risk profile allows the department to allocate resources effectively, not only in terms of number of patrols but also based on expertise.

Image: Chantele Dubois, “ The future of AI and predictive Policing” (2017)

Recommendation in a social perspective

The amount of data and the development of predictive analysis opens a door to optimize decision making that goes beyond crime prevention based on law enforcement perspective.

In the short term, mapping can further help allocation resources: decision about police stations, awareness campaigns and human capital allocation. In the mid-long term, mapping can be the initial step in targeted public actions to reduce criminality that goes beyond law enforcement. They can help other departments to create education and social programs to help reduce criminality by changing the region social reality.

Challenges moving forward

The advances in predictive analysis related to public security have opened a series of discussions about bias and discrimination based on income, race and other factors related to inequality issues in the country. The machine learning process utilizes historical data to make decisions and predictions, therefore it presents limitations in foreseeing structural changes in societies.

Questions remaining

Given the new trend of predictive, what are the social consequences that law enforcement predictive models can generate? how the limitations of the predictive model can be minimized in cities that are going through a series of structural changes (migration, increase population density, high economic growth)?

(Word count: 781)


[1] Tami Lin, Malgorzata Rejniak, “Smarter New York City: How City Agencies Innovate” Columbia University Press (2018)

[2] Cold Call Podcast, “ShotSpotter: A gunfire detection business looks for a new market” (2017), accessed in November 2018

[3] Rocco Parascandola, “EXCLUSIVE: NYPD ShotSpotter gunfire sensors improve rates of 911 calls, arrests” (2017), accessed in November 2018

[4] New York Police Department, “Report of the Finance Division on the Fiscal 2019 Preliminary Budget and the Fiscal 2018 Preliminary Mayor’s Management Report for the New York Police Department” Finance Division (2018)

[5] New York Police Department, “The way moving forward” NYPD Publication (2016)

[6] Chantele Dubois, “ The future of AI and predictive Policing” (2017), accessed in November 2018


Is There Value in Predicting When and How An Equipment Will Fail? Royal Dutch Shell Thinks So!


SUALAB: Deep Learning For Flexible Operation In Traditional Manufacturing

Student comments on Improving Public Safety with Machine Learning

  1. This is a very interesting take on how machine learning is being used for crime detection through tools like ShotSpotter.
    There is also another arena of crime prevention where machine learning has made some advances. For instance, a Chinese company developed Hikvision, a camera that can scan for suspicious anomalies like unattended bags at crowded venues, claiming to achieve 99% accuracy with advanced visual analytics applications.

    Reference read:

  2. Thank you for sharing this well-researched piece – very interested in how machine learning can be applied to issues in the public sector. I share your concerns about the inherent biases tied to these types of machine learning processes. Given the reliance on historical data, I worry these “heat map” predictive analytics will only perpetuate current racist policing practices among NYPD. Additionally, the placement of more acoustic sensors in predominantly low-income minority neighborhoods in New York City only serves to magnify the existing “over-policing” of these areas that result in more unjust arrests.

  3. Thank you for writing this piece. I’ve been personally following Shotspotter’s efforts for a number of years and I’m also equally fascinated by some of the questions that you posed towards the end of your article. On the question of Shotspotter’s social consequences, I believe that this might be an example of how the use of big data surveillance is currently shifting the the power balance between citizens and police, a shift that may in fact erode and not improve community trust.

    In NY and other metropolitan areas in the US, unreported gun shots are often a symptom of underlying tensions and mistrust held between community members and the police. The use of Shotspotter by police departments may be interpreted by some communities as the further encroachment of law enforcement into their vulnerable communities and could exacerbate and not alleviate the challenges faced by investigators and police officers within these communities.

  4. Thank you for sharing this! I was totally unaware about machine learning developments towards fighting crime, and I must say: I find its predictive efforts terrifying!

    An article on from barely a month ago goes:
    “Now called ShotSpotter Missions, HunchLab’s program applies historical crime data, as well as current and future indicators, to statistical models and machine learning to forecast crime at specific places and times. […] offers guidance on when to patrol each cell, what sort of crime is likely to occur there and what patrol tactics to use. […]”

    The system appears to be set up to support the status quo, and I strongly fear that being fed with historical data, this will reinforce the oppression that some minority communities already have to face.

    Also, here’s a critical article on its deployment across the river:

  5. Great read! Besides the bias and discrimination that you mention, another concern is that safety can be used as an excuse to overexert control on the public. For instance, machine vision has developed to a point whereby it is feasible to monitor where each specific individual is located at all times. While this seems like a scene out of a dystopian movie, we need to be careful that we don’t move towards that direction under the premise of creating a safer society.

Leave a comment