Optimizing Chicago’s Food Inspection Process Using Predictive Analytics

City:

Chicago, IL, USA

Organization:

City’s Departments: Chicago Department of Innovation and Technology (DoIT), Department of Public Health; 

External Partners: Civic Consulting Alliance and Allstate Insurance.

Project Start Date:

 June 2014 (Inferred, not specifically mentioned)

Project End Date:

 October 2014

Reference:

Schenk, T., Jr., Leynes, G., Solanki, A., Collins, S., Smart, G., Albright, B., & Crippin, D. (2015). Forecasting restaurants with critical violations in Chicago (pp. 1-13, Rep.). Chicago, IL: City of Chicago.

Hastie, T., & Qian, J. (2014, June 26). Glmnet Vignette. Retrieved October 6, 2017, from https://web.stanford.edu/~hastie/glmnet/glmnet_alpha.html 

Problem:

The Chicago Department of Public Health annually inspect over 15,000 restaurants with fewer than 3 dozen inspectors. The conventional inspection procedure is ineffective in prioritizing the inspections related to critical violations, increasing likelihood of spreading foodborne diseases to the public.

Technical Solution:

Combining past inspection results, weather, surrounding conditions, and business license data, the predictive model calculates the likelihood of a restaurant resulting in critical violations when inspected. The model calculates the number of failures (i.e., critical violations) through binomial and categorical answers. 

The rationale of the project is to calculates the % of inspections with violations under both the conventional operation and the model in month 1 over a 2-month pilot, and compare the values to interpret model efficiency.

Datasets Used:

• Dataset 1: Food Inspections, City of Chicago’s Data Portal, 2011.01-2014.01, 2014.09-2014.10
• Dataset 2: Business Licenses, City of Chicago’s Data Portal, 2011.01-2014.10, 2014.09-2014.10
• Dataset 3: Daily high temperature/weather information, forecast.io, 2011.01-2014.01, 2014.09-2014.10

Outcome:

Conventional operation: 15.8% of all conducted inspections yielded violations, of which 55% were found in month 1, and 45% in month 2. 

Model: 69% of violations were found in month 1, and 31% in month 2. The rate of finding violations increased by 25% in month 1, and 26.2% throughout the pilot. Violations were found on average 7.44 days earlier.

Issues that arose:

1. The pilot assumes that finding violations is time invariant, for which lacks a method to test.

2. The factor of inspector estimates is impractical to be included in a daily model, as it is difficult to ensure accuracy of the data collected. 

3.Additional data could have been included to supplement the model, such as online restaurant reviews. 

Status:

Operational

Entered by:

Claire (Kunyuan) Zhang 999542507  

kunyuan.zhang@mail.utoronto.ca

October 23, 2017