Overview
The Ames Housing Dataset, created by Dean De Cock, contains 81 features describing different aspects of residential homes in Ames, Iowa. The goal is to predict the Sale Price of each house.
This project covers the full modeling pipeline: exploratory data analysis, feature engineering, handling missing data and outliers, checking regression assumptions, and fitting multiple regression models.
Why This Approach?
The dataset’s complexity — 81 features with missing values, skewed distributions, and potential multicollinearity — demands a systematic approach. The strategy emphasizes:
- Thorough EDA to uncover relationships and guide feature selection
- Robust preprocessing to handle missing data and outliers while preserving data integrity
- Assumption checking before fitting any parametric model
Exploratory Data Analysis
The target variable SalePrice is right-skewed. A log transformation brings it closer to normality, which is important for linear regression’s residual assumptions.
import numpy as np
import matplotlib.pyplot as plt
plt.figure(figsize=(10, 4))
plt.subplot(1, 2, 1)
df['SalePrice'].hist(bins=50)
plt.title('SalePrice (original)')
plt.subplot(1, 2, 2)
np.log1p(df['SalePrice']).hist(bins=50)
plt.title('SalePrice (log-transformed)')
plt.tight_layout()
Handling Missing Data
Missing values were categorized by type:
- Structural missingness (e.g.,
PoolQCis NaN because there is no pool) → filled with"None" - Numeric missingness → filled with median
- High-missingness columns (>80% missing) → dropped
Model Performance Summary
| Model | RMSE (CV) | R² |
|---|---|---|
| Linear Regression | 0.1312 | 0.891 |
| Ridge Regression | 0.1287 | 0.894 |
| Lasso Regression | 0.1301 | 0.892 |
Ridge regression performed best, confirming that mild regularization helps with the multicollinearity present in this dataset.
Conclusion
Feature engineering and preprocessing contributed more to model performance than the choice of algorithm. The log-transformed target, careful imputation, and removal of high-leverage outliers were the most impactful steps. Ridge regression provided the best balance of bias and variance for this dataset.