netzer97/kc-house-price-regression

House Price Modeling – Regression & Classification (King County)

Presentation Video

Video (4–6 minutes): https://drive.google.com/file/d/1oWxT0g7jjDvenX_WFa0v7emERj_K2phe/view?usp=sharing

Project Overview

This project predicts residential house prices in King County, Washington using supervised machine learning. Two related tasks were addressed:

• Regression: Predicting the (log-transformed) continuous house price
• Classification: Predicting whether a house price is below or at/above the median

The workflow emphasizes exploratory analysis, feature engineering, model comparison, and evaluation, while carefully avoiding data leakage throughout the pipeline.

---

Dataset

The dataset contains information about residential properties sold in King County.

• Observations: 20,000+
• Features: Numerical and categorical attributes related to property size, location, condition, and amenities
• Regression target: House price (modeled as log(price) for improved stability)
• Classification target: Binary label (below / at-or-above median)

---

Exploratory Data Analysis (EDA)

EDA was used to understand distributions and major drivers of price.

Key findings:

• Strong relationship between price and size-related features (e.g., living area)
• Location-based variables significantly impact price
• Price distribution is highly right-skewed → motivates log transformations

Price distribution:

---

Feature Engineering

Several engineered features were created to improve model performance:

• House age (from yr_built)
• Renovation flag
• Living-to-lot ratio
• Total area (above + basement)
• Luxury score combining grade, view, and waterfront signal
• Log transforms for skewed numeric variables

Preventing data leakage

• Zipcode target encoding was computed using training data only, and then applied to test using a global-mean fallback for unseen zipcodes.
• For classification, the median threshold was computed on the training set only.

---

Regression Models

The following regression models were trained and compared:

• Linear Regression (baseline)
• Ridge Regression
• Lasso Regression

Selected model (Winner): Ridge Regression — chosen for strong predictive performance and stability from regularization.

Predicted vs. Actual (Ridge):

Saved artifact:

• winning_regression_model.pkl

---

Classification Models

The regression target was converted into a binary classification problem using a median split (train-only median).

The following models were trained and compared:

• Logistic Regression
• Random Forest Classifier
• Gradient Boosting Classifier

Evaluation used:

• Precision / Recall / F1-score
• Confusion Matrix

Selected model (Winner): Random Forest Classifier — best overall performance with fewer critical errors.

Confusion Matrix (selected model):

Saved artifact:

• winning_classification_model.pkl

---

Repository Contents

• winning_regression_model.pkl – final regression model
• winning_classification_model.pkl – final classification model
• Project notebook (.ipynb)
• README.md