CO₂ Car Emissions Prediction

🏗️ Project Context

This was a research project developed for the second term of the Post-Degree Diploma in Data Analytics in Langara College.

📌 Project Overview

This project builds linear regression models in R to predict car CO₂ emissions (grams/km) using the Government of Canada’s Fuel Consumption dataset. The analysis includes descriptive statistics, feature evaluation, and regression modeling to identify the best predictors of emissions and compare model performance.

🎯 Objective

• Analyze fuel consumption, engine size, cylinders, transmission type, and fuel type to understand their relationship with CO₂ emissions.
• Develop regression models to predict emissions and assess their accuracy.
• Compare stepwise, forward, backward, and subset selection methods to optimize the model.

🛠 Tools & Technologies

• Language: R
• Libraries: MASS, leaps, ggplot2, dplyr
• Techniques: Stepwise regression, subset selection, interaction effects, model validation

📊 Key Steps

1. Data Preparation

   • Cleaned and filtered the Government of Canada dataset.
   • Split into training (80%) and testing (20%) sets.

2. Descriptive Analysis

   • Explored numerical (engine size, cylinders, consumption) and categorical (fuel type, transmission, class, gears) variables.
   • Identified strong correlations (e.g., CO₂ vs fuel consumption: 0.935).

3. Modeling

   • Built multiple regression models (stepwise, forward, backward, exhaustive search).
   • Tested in total 9 models with and without interaction terms.
   • Compared Model number 8 (full model) vs. Model number 9 (reduced model).

4. Evaluation

   • Achieved high predictive accuracy (Adjusted R² ≈ 0.9992).
   • Residual analysis confirmed improved variance stability after including interaction terms.

🚀 Results

• The full model, which included consumption, fuel type, year, gears, transmission, engine, and the interaction term (consumption × fuel), emerged as the strongest predictor of CO₂ emissions (RMSE = 1.83, R² ≈ 0.9992).
• The reduced model, using only consumption, fuel type, and their interaction, offered a simpler structure while still maintaining high accuracy (RMSE = 2.47, R² = 0.9985), making it the most effective for practical use.
• Demonstrated that smaller models can still deliver reliable predictive performance.

📂 Repository Structure

├── Data/              # Raw and cleaned datasets (train/test splits)
├── Src/           # R scripts for analysis and modelling
├── Documentation/f   # Full project report
└── README.md          # Project documentation