This project is part of the Ironhack Data Science & Machine Learning Bootcamp and follows a Kaggle-style competition format. The dataset covers house sales in King County, Washington (including Seattle) between May 2014 and May 2015.
The primary objective is to analyze housing prices and identify the key factors that influence them, with a special focus on properties priced above $650,000. Beyond exploratory analysis, we also apply machine learning models to predict housing prices more accurately.
.
├── Mini_Project.ipynb # Jupyter notebook with full analysis and code
├── data/ # Raw and processed datasets
├── images/ # Visualizations exported from notebook
├── IronKeggle - Project Presentation.pdf # Project presentation
└── README.md # Project documentation-
Perform Exploratory Data Analysis (EDA) to understand property characteristics.
-
Engineer new, more meaningful features (e.g., house age, renovation flag, price per m²).
-
Identify key drivers of housing prices (size, rooms, location, etc.).
-
Handle missing values, duplicates, and outliers. Detect and handle outliers, missing values, and categorical variables.
-
Build and tune machine learning models for price prediction.
-
Communicate insights through data visualization and clear reporting.
-
Interpret and rank feature importance to highlight the most influential property characteristics.
- Categorical vs. numerical feature classification.
- Converted yr_built → age and yr_renovated → renovated (binary).
- Converted date to datetime type for temporal analysis.
- Added age of the property.
- Created renovation flag.
- Built geolocation clusters using KMeans on latitude and longitude.
- Generated price per square meter for relative valuation.
- Removed outliers using IQR filtering.
- Handled categorical variables: - Ordinal → treated as ordered (e.g., condition, grade). - Nominal → one-hot encoded.
- Normalized continuous features with RobustScaler.
- Applied log transformation to target (price) to reduce skew.
-
Larger living area strongly correlates with higher price.
-
Location (lat/long clusters) is a key driver of housing values.
-
Renovated properties and higher-grade houses sell for a significant premium.
-
Price distribution is right-skewed, requiring transformation.
-
Baseline models tested.
-
XGBoost Regressor with hyperparameter tuning via GridSearchCV (5-fold CV) provided the best balance of accuracy and generalization.
-
Feature importance analysis showed:
- Location (lat, long, and cluster feature) as the strongest driver. - Size (sqft_living, sqft_lot, sqft_living15) as major contributors. - Age, bathrooms, bedrooms, and grade also highly influential. - Renovation status positively impacts predicted price. -
Discussion of predictive power and limitations.
-
Location matters most: longitude, latitude, and neighborhood clusters dominate the model.
-
House size (living area, lot size, nearby house sizes) is a critical factor in pricing.
-
Renovations & age influence property value significantly.
-
Feature engineering (clusters, age, log-price) improved model robustness and reduced reliance on a single dominant feature.
- Add external datasets (schools, crime rate, amenities) for richer context.
- Experiment with ensemble methods (Stacking, LightGBM, CatBoost).
- Improve predictive models and try new ones.
- Deploy a streamlit app or dashboard for interactive price prediction.
- Perform geospatial visualization to highlight neighborhood trends.
Python 🐍 Pandas, NumPy Matplotlib, Seaborn Scikit-learn XGBoost
Hi, I’m Katherine! 👋
I’m currently completing the Ironhack Data Science Bootcamp and building projects in Machine Learning & Data Analysis.
My long-term goal is to become a Machine Learning Engineer and specialize in real estate and financial applications of AI.
📧 Contact me: pktorian@gmail.com
🔗 LinkedIn | GitHub