Skip to content

Latest commit

 

History

History
175 lines (111 loc) · 7.04 KB

README.md

File metadata and controls

175 lines (111 loc) · 7.04 KB
Raw

Chicago, IL : Car Crash Analysis & Predictive Modeling

Predicting severity of car crashes with Machine Learning Models

Authors:

Christos Maglaras, Marcos Panyagua, Jamie Dowat

Date: 3/12/2021

chicago

Contents

  • notebooks: Contains personal contributor notebooks
  • final_notebook.ipynb : Contains more detailed analysis and walks through modeling process.
  • src : Contains source code for final_notebook.ipynb.
  • data : Contains csv files used for modeling (See below for description and links).
  • presentation.pdf : Slide deck for non-technical presentation

Data: Chicago City Data Portal

ccdp

1. Crashes:

  • Number of Rows: 482,866

  • Shows crash data from crash from the Chicago Police Department's E-Crash system

  • "All crashes are recorded as per the format specified in the Traffic Crash Report, SR1050, of the Illinois Department of Transportation."

2. People:

  • Number of Rows: 1,068,637

  • Information about people involved in a crash and if any injuries were sustained.

3. Vehicles:

  • Number of Rows: 987,148

  • Information about vehicles ("units") involved in a traffic crash.

Stakeholder: Chicago Department of Transportation

Click here to check out their website!

cdot

Business Understanding

Just a week ago, the National Security Coundil released a report containing some disturbing statistics from 2020. The first paragraph begins as follows:

"For the first time since 2007, preliminary data from the National Safety Council shows that as many as 42,060 people are estimated to have died in motor vehicle crashes in 2020. That marks an 8% increase over 2019 in a year where people drove significantly less frequently because of the pandemic."

According to their data, the US hasn't seen an increase like this since 1924.

Following this trend, the Governor's Highway Safety Association reported that pedestrian fatality rate has reached a 30-year high, with nighttime pedestrian fatalities having increased by 67%, and a 16% increase in daytime fatalities, highlighting the need for safer road crossings and increased efforts to make pedestrians and vehicles more visible.

Narrowing our focus even further, in Illinois, around 1000 people were KILLED in motor vehicle crashes in 2019 alone.

Advocates for Highway and Auto Safety have scored all US states against their Roadmap for State Highway Safety Laws, a set of 16 laws that cover occupant protection (selt belt, helmet laws), child protection, and teen driving.

scoring

When Illinois is scored against this Roadmap, it has been given a yellow rating (Caution), since it still lacks the following safety laws:

  • All-Rider Motorcycle Helmet Law
  • Booster Seat Law
  • GDL (Graduated Driver's License) – Minimum Age 16 for Learner’s Permit
  • GDL – Stronger Nighttime Restriction Provision
  • GDL- Stronger Passenger Restriction Provision
  • GDL- Age 18 for Unrestricted License

Currently, the Chicago Department of Transportation is working with the city's new initiative, Vision Zero, to reduce accidents on the road. In Vision Zero's report and action plan, they used crash data to identify high crash corridors in the city as well as other important trends to guide education, road safety improvements, and more.

Bearing all of this in mind, we had these current safety movements and road safety problems guide our exploration and modeling of this data.

Even more importantly, this business understanding was the sole influencer of our target choice, severity of crash based on injury.

Predictive Modeling Preview

In terms of business problems, we found one of the ways a predictive model could most help with this business problem is to determine how different factors of the crash determine the severity of injuries in the crash.

We compared results from Logistic Regression, K-Nearest Neighbors, Decision Tree, Naive Bayes, and Random Forest Classifier models, before settling on our final model produced by a Bayesian-Optimized XGBoost Classifier.

We experimented with both a BINARY classification and a TERNARY, with our final model using as TERNARY classification. The targets are defined as follows, using the MOST_SEVERE_INJURY column in the Crashes dataset:

  • BINARY:

    • Class 0: No Injury
    • Class 1: Injury

  • TERNARY

    • Class 0: No Injury
    • Class 1: NON-INCAPACITATING injuries
    • Class 2: INCAPACITATING or FATAL injuries

EDA

topcrashsites

driveraction

Showcasing Class Imbalance

classimbalance

Major Modeling Takeaways

1st Priority Metric: RECALL

We focused first and foremost on improve RECALL, since the most damaging false predictions would be false negatives (predicting no injury, when there was actually injury).

First Simple Model

Recall: 22%

fsm

FINAL MODEL

Recall: 85%

test

Top 5 Features - Final Model (by Gain Metric)

  • UNIT_TYPE - DISABLED VEHICLE (810)
  • DRIVER_ACTION - DISTRACTION - FROM OUTSIDE VEHICLE (181)
  • PERSON_TYPE - DRIVER (11)
  • VEHICLE_DEFECT - ENGINE/MOTOR (8.71)
  • VEHICLE_DEFECT - WHEELS (8.4)

To access all of our sources, or for more details about our process, see final_notebook.ipynb or presentation.pdf.

Repository Structure

├── data 
....├── traffic_crashes_chicago.csv
....├── traffic_crashes_people.csv
....└── traffic_crashes_vehicles.csv
├── img
....├── cdot.png
....├── chicago_night_drive.jpg
....├── chicagocitydataportal.jpg
....├── finalmodeleltestcm.png
....├── finalmodeltraincm.png
....├── scoringsafety.png
....└── visionzeroquotes.png
├── notebooks
....├── Christos_notebook.ipynb
....├── jamie_notebook.ipynb
....└── marcos_eda.ipynb
├── src
....├── __init__.py
....├── data_cleaning.py
....├── eda.py
....└── models.py
├──presentation.pdf
├── final_notebook.ipynb
└── README.md