National_Collision_DB_Group407
The data we are using is from the National Collision Database, it can be found on the Government of Canada website. It is a database that contains all police-reported motor vehicle collisions on public roads in Canada. The data is from 1999 to 2017. Each row provides several data points for a passenger with the detailed summary statistics of the collision. Due to the huge amount of data in this dataset, we are only using the most recent data from 2017.
Research Question: What features are indicative of a accident causing fatality?
We started our project by first wrangling and cleaning our dataset. In the wrangling stage, we will limit the features to case level variables, which includes road conditions, time of accident, weather etc. We restricted other personal and vehicle level features as we are mainly interested in external causes of fatal accidents that can be controlled with intervention.
The C_ISEV feature (Collision Severity) provides the fatality status of the accident which will be our target. The targets have two classes: Collision producing at least one fatality and Collision producing non-fatal injury. Our goal is to come up with a model that is able to perform binary classification between the two classes of fatality status given the following features:
After splitting the data into train and test sets, we would like to predict which features strongly predict the fatality. While performing exploratory data analysis we found our data is highly imbalanced between the two target classes. There are 179,714 cases of non-fatal collisions and 2,559 cases of fatal collisions which may lead to biased accuracy of model evaluation.
We therefore took the extra step of balancing our data by downsampling the majority group. We choose to use more robust models like random forest and logistic regression for our classification problem. These models scales well to our large dataset with good performance. After selecting the most predictive features and tuning hyper-parameters of our models, we ended up with reasonable accuracy in predicting fatal accidents. This would be useful for road assistant and medical agencies to prepare for emergencies under certain conditions that are more prone to having severe and fatal accidents.
More information about the meaning of each class and variables can be found here
Usage
To run the download script file, clone this GitHub repository, install the dependencies listed below, and run the following commands at the command line/terminal from the root directory of this project:
Rscript src/data_read.R --filepath='https://opendatatc.blob.core.windows.net/opendatatc/NCDB_2017.csv'
python src/data_clean.py --read_path= data/file.csv --write_path= data/
python src/eda.py --read_path= data/cleaned_train_data.csv --write_path= results/
Rscript -e "rmarkdown::render('doc/eda_report.Rmd')"
python src/ml_lgr_rf.py
Rscript -e "rmarkdown::render('doc/final_report.Rmd')"
Dependencies
Python 3.7.3 and Python packages:
- docopt==0.6.2
- pandas==0.25.1
- matplotlib==3.1.1
- altair==3.2.0
- numpy==1.17.2
- sklearn==0.0
- pandas-profiling==2.4.0
License
National Collision Database documents are licensed under the Open Government Licence - Canada. If re-using/re-mixing please provide attribution and link to this webpage.
References
Transport Canada. 2017. “National Collision Database.” Government of Canada; https://open.canada.ca/data/en/dataset/1eb9eba7-71d1-4b30-9fb1-30cbdab7e63a.