Designing an Explainable Predictive Policing Model to Forecast Police Workforce Distribution in Cities

Despite extensive research, measurable benefits of predictive policing are scarce. We argue that powerful models might not always help the work of officers. Furthermore, developed models are often unexplainable, leading to trust issues between police intuition and machine-made prediction. We use a j...

Full description

Saved in:  
Bibliographic Details
Authors: Parent, Mark (Author) ; Tremblay, Sébastien (Author) ; Deslauriers-Varin, Nadine (Author) ; Falk, Tiago H. (Author) ; Gagnon, Claudele (Author) ; Lemaire, Noémie (Author) ; Roy, Aurélien (Author)
Format: Electronic Article
Language:English
Published: 2020
In: Canadian journal of criminology and criminal justice
Year: 2020, Volume: 62, Issue: 4, Pages: 52-76
Online Access: Volltext (Resolving-System)
Journals Online & Print:
Drawer...
Check availability: HBZ Gateway
Keywords:
Cognitive Psychology
Data science
Human-computer interactions
Neuroergonomics
Neurophysiology
Description
Summary:Despite extensive research, measurable benefits of predictive policing are scarce. We argue that powerful models might not always help the work of officers. Furthermore, developed models are often unexplainable, leading to trust issues between police intuition and machine-made prediction. We use a joint approach, mixing criminology and data science knowledge, to design an explainable predictive policing model. The proposed model (a set of explainable decision trees) can predict police resource requirement across the city and explain this prediction based on human-understandable cues (i.e., past event information, weather, and socio-demographic information). The explainable decision tree is then compared to a non-explainable model (i.e., a neural network) to compare performance. Analyzing the decision tree behaviour revealed multiple relations with established criminology knowledge. Weather and recent event distribution were found to be the most useful predictors of police workforce resource. Despite wide research showing relationships between socio-demographic information and police activity, socio-demographic information did not contribute much to the model’s performance. Though there is a lack of research on measurable effects of predictive policing applications, we argue that combining human instinct with machine prediction reduces risks of human knowledge loss, machine bias, and lack of confidence in the system.
ISSN:1911-0219
DOI:10.3138/cjccj.2020-0011