🚗 Road Accident Severity Prediction

A machine learning techniques to analyze road accidents in the US and predict their severity. This project includes data preprocessing, natural language processing (NLP) on accident descriptions, and training a machine learning model to classify accident severity.

---

📁 Repository Structure

• ML_Engineer_Task.ipynb
  Preprocesses raw accident data by handling missing values, feature engineering, and scaling. Outputs a processed dataset in .csv or .pkl format.

• ML_Engineer_Task_Colab.ipynb
  Loads the preprocessed data and applies NLP techniques to accident descriptions. Trains a Random Forest Classifier with SMOTE and outputs a trained model (Random_Forest_Classifier_SMOTE.pkl).

---

📊 Dataset Attributes

The dataset consists of the following attributes, each providing crucial information about road accidents. Here's a detailed description:

Attribute	Description	Nullable
ID	Unique identifier for each accident record.	No
Severity	Severity level of the accident (1 to 4), where 1 indicates minor impact and 4 indicates significant impact.	No
Start_Time	Start time of the accident in the local timezone.	No
End_Time	End time of the accident's impact on traffic flow.	No
Start_Lat	Latitude of the starting point of the accident.	No
Start_Lng	Longitude of the starting point of the accident.	No
End_Lat	Latitude of the ending point of the accident.	Yes
End_Lng	Longitude of the ending point of the accident.	Yes
Distance(mi)	Length of the road segment affected by the accident.	No
Description	Natural language description of the accident.	No
Number	Street number in the address field.	Yes
Street	Street name in the address field.	Yes
Side	Side of the street (Right/Left).	Yes
City	City name.	Yes
County	County name.	Yes
State	State name.	Yes
Zipcode	Zip code of the location.	Yes
Country	Country name (e.g., US).	Yes
Timezone	Timezone of the accident's location (e.g., Eastern, Central).	Yes
Airport_Code	Nearest airport-based weather station to the accident location.	Yes
Weather_Timestamp	Timestamp of the weather observation for the accident.	Yes
Temperature(F)	Temperature at the time of the accident (in Fahrenheit).	Yes
Wind_Chill(F)	Wind chill at the time of the accident (in Fahrenheit).	Yes
Humidity(%)	Humidity percentage at the time of the accident.	Yes
Pressure(in)	Atmospheric pressure at the time of the accident (in inches).	Yes
Visibility(mi)	Visibility at the time of the accident (in miles).	Yes
Wind_Direction	Wind direction at the time of the accident.	Yes
Wind_Speed(mph)	Wind speed at the time of the accident (in miles per hour).	Yes
Precipitation(in)	Precipitation amount at the time of the accident (in inches).	Yes
Weather_Condition	Weather conditions (e.g., Rain, Snow, Fog).	Yes
Amenity	Indicates the presence of nearby amenities.	No
Bump	Indicates the presence of nearby speed bumps.	No
Crossing	Indicates the presence of a nearby crossing.	No
Give_Way	Indicates the presence of a nearby "Give Way" sign.	No
Junction	Indicates the presence of a nearby junction.	No
No_Exit	Indicates the presence of a nearby "No Exit" sign.	No
Railway	Indicates the presence of nearby railways.	No
Roundabout	Indicates the presence of a nearby roundabout.	No
Station	Indicates the presence of a nearby station.	No
Stop	Indicates the presence of a nearby stop sign.	No
Traffic_Calming	Indicates the presence of nearby traffic-calming measures.	No
Traffic_Signal	Indicates the presence of nearby traffic signals.	No
Turning_Loop	Indicates the presence of a nearby turning loop.	No
Sunrise_Sunset	Period of the day (Day/Night) based on sunrise and sunset.	Yes
Civil_Twilight	Period of the day based on civil twilight.	Yes
Nautical_Twilight	Period of the day based on nautical twilight.	Yes
Astronomical_Twilight	Period of the day based on astronomical twilight.	Yes

---

🛠️ Features

✅ Data Preprocessing

• Handles missing values using group-based imputations.
• Groups Weather_Condition categories into broader groups (e.g., Rain, Snow, Fog).
• Encodes categorical variables and scales numerical features.

📊 Natural Language Processing

• Cleans, tokenizes, and lemmatizes accident descriptions.
• Vectorizes text data using TfidfVectorizer for feature extraction.

🧠 Model Training

• Uses a Random Forest Classifier.
• Employs SMOTE to handle imbalanced class distribution.
• Outputs a model file (Random_Forest_Classifier_SMOTE.pkl) for future use.

---

🚀 Getting Started

1️⃣ Preprocess Data

1. Clone the repository:

   git clone https://github.com/your-username/your-repo-name.git

2. Open and run the notebook ML_Engineer_Task.ipynb:
   • Loads the raw dataset.
   • Cleans and preprocesses data.
   • Saves the output as accident_data.pkl or accident_data.csv.

2️⃣ Train the Model

1. Open ML_Engineer_Task_Colab.ipynb in Google Colab or your preferred environment.
2. Load the preprocessed file (.pkl or .csv).
3. Run the notebook to:
   • Perform NLP on accident descriptions.
   • Train the model and save it as Random_Forest_Classifier_SMOTE.pkl.

---

📦 Outputs

• Processed Dataset: accident_data.pkl or accident_data.csv
• Trained Model: Random_Forest_Classifier_SMOTE.pkl

---

🛑 Prerequisites

Install the required Python libraries:

pip install pandas scikit-learn nltk matplotlib seaborn

---

📊 Results

• Validation Accuracy: ~96.9%
• Class Imbalance Handling: SMOTE improves performance on underrepresented severity classes.

---

📅 Future Enhancements

• Experiment with other ML algorithms like Gradient Boosting or neural networks.
• Integrate real-time accident data for dynamic predictions.
• Enhance feature engineering for improved accuracy.

---

🤝 Contribution

Contributions are welcome! If you have ideas for improvement:

1. Fork this repository.
2. Create a new branch (feature-branch-name).
3. Commit changes and push.
4. Open a Pull Request.

---