Skip to content

Latest commit

 

History

History
79 lines (56 loc) · 2.93 KB

writeup.md

File metadata and controls

79 lines (56 loc) · 2.93 KB
Raw

Hit and runs in Chicago

Ryan Solava

Abstract

In this project, we use crash data from the City of Chicago to train classifiers to predict if a crash is likely to be a hit and run. A variety of models were trained and compared

Design

Hit and runs are on the rise in Chicago. The victims of hit and runs are double victims in that they must deal with the affects of the crash, but also bare the financial burdens themselves without another party to hold responsible. As part of the City of Chicago's response to this situation, here in this project we use past crash data to predict if a given context is likely to produce a hit and run. This will be useful to city planners and police, in helping them to best allocate resources that could prevent these occurrences.

Data

  • Chicago crash data Data on all crashes in Chicago from 2017-2021 are publicly available on the city's data portal. There were about 581,000 crashes. Available features include
  • Time and date of crash
  • Longitude and latitude
  • Lighting
  • Weather conditions
  • Road conditions
  • Road type
  • Traffic signal type
  • Whether it was in a work zone

And many others. There are several features such as the crash type, amount of damage, number and seriousness of injuries that are about the results of the crash. This data was used for some initial analysis, but was not used to train the model, since it could not be know until after a crash occurs. Our target, whether or not a crash was a hit or run, is also available in the dataset.

Algorithms

The following soft classifiers were trained:

  • Logistic regression
  • K-nearest neighbors
  • Decision tree
  • Random forest
  • Extremely random tree
  • XGBoost

Naive Bayes was also considered, but since there is a mix of continuous and binary features, it was difficult to get an effective model.

Of these models, XGBoost was the best performing with an area under the receiver operator curve of .655. This was after hyperparameter tuning.

Tools

  • XGBoost* for the XGBoost classifier
  • scikitlearn was where all the other classifiers came from, I also used their model selection package to tune hyperparameters.
  • Geopandas to plot and display geographic data.
  • Sweet Viz for initial EDA.

Communication

I primarily communicated my results through slides and presentation. Also, my code and visualizations are posted to my personal GitHub.