August 14, 2018
Fire Risk Analysis: Predicting Fire Risk to Prioritize Commercial Property Fire Inspections
The Fire Risk Analysis project uses fire incident and property data to develop predictive models of structure fire risk in partnership with the Pittsburgh Bureau of Fire (PBF) and the Department of Innovation and Performance (I&P). PBF conducts regular fire inspections of commercial properties, as stipulated by the municipal fire code. This project helps PBF prioritize their property inspections with data-driven insights from the fire risk analyses from machine learning models, implemented in a data dashboard and interactive map visualization, in order to target their inspections at the properties at greatest risk of fire.
STATUS: Our team has developed and deployed a statistical model using “machine-learning” to predict structure fire risk in properties around Pittsburgh. We developed (or, “trained”) the model using data from historical fire incident data from PBF, property inspection data from the Department of Permits, Licenses, and Inspections, and property assessment data from the Allegheny County property assessment office. We evaluated multiple model types (e.g. random forests, SVM, XGBoost), and tuned these models to find the best-performing model and generate risk scores for each commercial property in the city, as a function of its probability for a fire to occur at that address. The risk scores generated by the model are displayed on a data dashboard and an interactive map developed by the Department of Innovation and Performance. These tools are used by fire inspectors and fire chiefs to inform the Bureau of Fire’s prioritization of property fire inspections, so they can inspect the properties at greatest risk of fire.
The model is currently deployed on their servers and retrains on a regular basis, as new data is generated. In the months since the first model was trained and risk scores generated, 14 of the 45 (31%) building-related fire incidents (not all of which are “working fires”) occurred in one of the medium or high-risk properties, significantly higher than the 0.20% base rate for fire incidents in the city. We are currently monitoring the stability of the model’s performance over time and conducting experiments using “neural network” models to better capture some of the temporal dynamics of incident events around the city. We have found that the model has remained quite stable over the 6 months since its official deployment in February, 2018, with standard deviations of our key performance metrics less than 0.01, over the 16 model iterations since February.
Our future work includes incorporating new data sets, experimenting with additional model types (e.g. recurrent neural networks, reinforcement learning, and/or “active learning”), and expanding this approach to predict fire risk in residential properties at the census block level. Our team is also working to make the model more easily usable by other cities’ fire agencies.
If you’re a student with data science or machine learning experience and are interested in contributing to this project, or if you’re a representative from a city interested in adapting this model, please contact the project lead below.
This work was recently accepted to the KDD conference:
Singh, B., Hu, Q., Chen, J., Chen, F., Lee, J., Kuo, N., Narang, P., Batts, J., Arnold, G., Madaio, M. (2018). A dynamic pipeline for spatio-temporal fire risk prediction. In Proceedings of the 2018 ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) (in press). [pre-print pdf]
Chief Jones, Lt. Jason Batts
Pittsburgh Bureau of Fire
Laura Meixell, Geoffrey Arnold
Pittsburgh Department of Innovation and Performance
Ph.D. Student, Human-Computer Interaction Institute
Carnegie Mellon University