House Price Prediction

This project builds a regression model to predict house sale prices from https://www.kaggle.com/competitions/house-prices-advanced-regression-techniques/data. The process covers exploratory data analysis (EDA), data preprocessing, model training, and model interpretation using SHAP values.

---

Features

• Numeric Features: GrLivArea, OverallQual, GarageCars, 1stFlrSF, YearBuilt, etc.
• Categorical Features: HouseStyle, Foundation, GarageType, etc.

Categorical variables are processed using OneHotEncoder to convert categories into binary numerical features. Numeric features are scaled using **StandardScaler`.

---

🛠️ Libraries & Tools

• pandas
• numpy
• scikit-learn
• xgboost
• seaborn
• matplotlib
• SHAP
• joblib
• scipy
• lightgbm

---

Workflow

1. Data Cleaning & Preprocessing

• Handling missing values:
  • Categorical columns: filled with 'None'
  • Numeric columns: filled with median or 0
• Log transformation applied to SalePrice to reduce skewness and improve model performance.

2. Exploratory Data Analysis (EDA)

• Distribution plots for target variable (before and after log transform)
• Scatterplots to check correlation between features and sale price
• Skewness check to ensure the target is close to normally distributed
• Log-transform SalePrice to fix skewness value

3. Model Building

Models used:

• Ridge Regression
• Lasso Regression
• Random Forest Regressor
• XGBoost Regressor

Each model was integrated into a Pipeline containing:

• Preprocessor (for scaling numeric features and encoding categoricals)
• Regressor (the model)

4. Hyperparameter Tuning

• Ridge Regression tuned using GridSearchCV to find the optimal alpha value.
• Cross-validation used to assess generalization performance.

5. Model Evaluation

Metrics:

• RMSE (Root Mean Squared Error)
• R² Score

Both evaluated on the log-transformed predictions and then converted back to the original scale.

6. Model Interpretation

• SHAP Summary Plot visualizes each feature's impact on model predictions.
• High SHAP values indicate features that push predictions higher or lower, with color showing whether the feature value was high or low.

7. Another model training

Training model using Light GBM Regression algorithm. The model result:

• RMSE: 28681.16
• R2 Train: 0.9882
• R2 Test: 0.8928

Tuning model with result:

• Test RMSE: 29933.29
• R2 Train: 0.9777
• R2 Test: 0.8832

---

📎 How to Run

1. Clone this repository:

   git clone https://github.com/RaymussenArthur/House-Price.git