| City: |
Atlanta, Georgia, United States of America
|
| Organization: |
- Atlanta Fire Rescue Department (AFRD)
- Georgia Institute of Technology
- Data Science for Social Program
|
| Project Start Date: |
Mid 2015
|
| Project End Date: |
Present
|
| Reference: |
- Madaio, M., Chen, S., Haimson, O., Zhang, W., Cheng, X., & Hinds-Aldrich, M. et al. (2016). Firebird. Proceedings Of The 22Nd ACM SIGKDD International Conference On Knowledge Discovery And Data Mining - KDD '16. http://dx.doi.org/10.1145/2939672.2939682.
- Predicting Fire Risk and Prioritizing Fire Inspections. (2017). Firebird. Retrieved 23 September 2017, from http://firebird.gatech.edu/
|
| Problem: |
Content
- To better move away from their traditional method of basing inspections on intuition, AFRD are challenged in prioritizing fire inspections, or identifying new properties requiring inspections given their limited resources.
Objective
- Develop a data-driven process that can assist the AFRD in managing their resource effective for fire hazards across Atlanta.
Deliverable
- Populate an interactive map with a visualized breakdown of fire incidents, property information and risk scores to help AFRD make informed decisions about fire inspections.
|
| Technical Solution: |
Supervised Learning: Evaluate and predict the accuracy of the model based on using fire incident data in July 2011 to March 2014.
- 58 independent variables were further used to help predict fire as an outcome variable for each property.
- Training Data: July 2011 - March 2014
- Test Data: April 2014 - March 2015
Based on a grid search with 10-fold cross validation, AFRD tested 4 models to determine the best model and parameters.
- Logistic Regression
- Gradient Boosting
- Support Vector Machine (SVM) - yielded the best results (71.36% True Positive Rate and 20% False Positive Rate)
- Random Forest - yielded the second best results (69.28% True Positive Rate and 20% False Positive Rate)
|
| Datasets Used: |
- Dataset 1: Fire Incident, Atlanta Fire Rescue Department, 2011-2015
- Dataset 2: Fire Permits, Atlanta Fire Rescue Department, 2012-2015
- Dataset 3: Parcel, City of Atlanta
- Dataset 4: Strategic Community Investigation, City of Atlanta
- Dataset 5: Business Licenses, City of Atlanta
- Dataset 6: Crime, Atlanta Police Department, 2014
- Dataset 7: Liquor Licenses, Atlanta Police Department, 2014
- Dataset 8: Neighbourhood Planning Unit, Atlanta Regional Commission
- Dataset 9: Demographics, U.S. Census Bureau
- Dataset 10: Socioeconomic, U.S. Census Bureau
- Dataset 11: CoStar Properties, CoStar Group. Inc
- Dataset 12: Google Place, Google Place APIs
|
| Outcome: |
Previous Inspection Performance Process
- Much of the previous inspection process fell onto fire inspectors to notice new properties and initiate an inspection while patrolling the area.
- There have been no previous method to formally prioritize their inspections or schedule their daily inspections.
Post Inspection Performance Process
- Through the use of the predictive model, AFRD found an additional 6,096 commercial properties to inspect annually
- It is unfeasible without significant changes in the organizational process, as such the initial effort was to assign the new 69 high-risk properties (27 out-of-date propeties, 13 required new permits, 15 were out of business)
- Increase communication between departments and stakeholders to further data exploration and resource allocation.
Technology Adoption
- Development and deployment of an interactive map-based visualization tool of fire incidents, currently inspected properties, and potentially inspected properties.
|
| Issues that arose: |
Data
- Clean and Useable Data: Despite the volume of data available, AFRD was not able to utilize the 20,000 commercial properties in the city given the lack of completeness in many of the dataset. As such, only 8,223 properties were used in their predictive fire risk mode, with only 5,022 having a risk score.
- Universal Format: Different organization did not share a universal database of buildings or identification numbering convention, as such there was more of a technological difficulty when joining various datasets. Elaborate process of fuzz text-matching and address verification was also required for building's dataset.
- Latest Versions: Access to the latest data was also a challenge for the AFRD to access given the lack of data sharing among city departments. For instances, the Office of Building's Business License datbase may update regularly; however, without a method for AFRD to access the data, the knowledge of business closing without AFRD knowledge can lead to a waste of resource.
Organizational Process
- There lacks an organizational procedure for AFRD to add properties to the list of regular inspection, or to determine their frequency of inspection. As such, there is no innovative method of updating properties status beyond having it redone on a regular basis.
Policy Standpoint
- With the current practice of the municipal Fire Code to regularly examine specific set of commercial property type, AFRD will need to further discuss on how to revise the code to determine the difference in inspection type, priority, and frequency due to the fire risk associated with property types.
- While the AFRD is geared towards commercial property, Atlanta's fires typically occur in residential properties. This will require additional rethinking of how the model can be applied in a residential community context.
Labour Resources
- With the increase in the number of properties to go up by 237%, there is currently no incentive in how the fire inspectors can efficiently inspect properties.
|
| Status: |
Operational/Improvement
Refine, Expand and Validate Predictive Model
- Identify missing and erroneous data sources
- Promote a stronger integration across the City's dataset
- Train the predictive models to utilize dataset more effectively by providing fewer information variables, but creating more applicable properties to investigate.
Additional Dynamic Sources of Data
- Prior evidence of violation of inspections
- Health and Wellness inspections
- Information of Occupancy
- Behaviour Sources
|
| Entered by: |
October 2nd, 2017: Alvin Fan, alvin.fan@mail.utoronto.ca
|