Built a web application using Streamlit and deployed on Heroku.

https://traffic-severity-prediction.herokuapp.com/

With the increasing of road traffic infrastructures, motor vehicles, drivers, and traffic flow, the role of road traffic in supporting and guiding economic and social development is becoming more and more obvious. As a result, road traffic safety has increasingly become a key issue in concerning the safety of people’s lives and property, as well as affecting the quality and efficiency of economic and social development. Road traffic accidents are the process of simultaneous damage of people or things, which caused by the coupling imbalance of dynamic and static factors such as human, vehicle, road, and environment. Therefore, it is necessary to study the influencing factors, as well as the classification and identification model of the severity of road traffic accident, so as to pave the way for improving the safety level of road traffic.

This is a multi-class classification problem where we are predicting the severity of accident :

• Slight Injury
• Fatal Injury
• Serious Injury

based on the other 31 features.

This data set is Road Traffic Accidents from Kaggle. The data set has been prepared from manual records of road traffic accident of the year 2017-20. All the sensitive information have been excluded during data encoding and finally it has 32 features and 12316 instances of the accident.

• Exploratory Data Analysis is the first step of understanding your data and acquiring domain knowledge.

• The dataset has around 16 features with missing values. This missing values are imputed using Predictive Imputation technique where I used the known values to predict the missing values.

• The Dataset was quite imbalanced with 10415 records with Slight injury, 1743 records withSerious injury anf just 158 records with Fatal injury.
• I used SMOTE, Random Under/Over Sampling, NearMiss method for balancing the dataset.

• On using Correlation and Chi-square method, I found that some features were moderately correlated so I removed the features with collinearity. Finally, we use 16 feature for training process.

• On training my model using several classification algorithms such as XGBoost, Random Forest, Extra Trees, Decision Trees, the model trained with XGBoost gave good results and less memory.
• Used KFold with 5 splits cross validation with hyper-parameter tuning on XGBoost Classifier (baseline model) using GridSearchCV.
• Also, I found that my baseline model (XgBoost Classifier) was overfitting the dataset. On investigation I found that the dataset was affected by Curse of Dimensionality. So I reduced the dimensions and trained my model again.
• After retraining my model, I found that it was generalizing well with an accuracy of 74%.
• As per the problem statement I used F1 Score as the evaluation metric for my model.

• Used SHAP based on test prediction.

• Clone this repository and check the requirements.txt:

  git clone https://github.com/sonnguyen129/Accident-Severity-Prediction
  cd Accident-Severity-Prediction
  pip install -r requirements.txt

• Simply run:

  streamlit run app.py