Case study: Feature Engineering-- Ames house price prediction

1 Problem statement

In this case study, you will prepare Ames Housing Dataset in a csv file in a way that it is suitable for a ML algorithm. You will achieve this by first exploring the data and performing feature transformations on provided dataset of house price prediction ML problem. You are required to train a ML model by using linear regression, ridge regression and lasso regression for predicting house prices.

2 Steps

• 2.1 Load data set
• 2.2 Exploratory Data Analysis (EDA)
• 1. Histograms

• 2. Heatmap

• 3. Scatterplots

• 4. Scatter matrix

• 5. Correlation between other features and 'SalePrice'

The target 'SalePrice' variable is highly correlated with features such as OverallQual, GrLivArea, GarageCars, GarageArea and TotalBsmtSF among others.

• 2.3 Process dataset for ML

Steps:

• 1. Handle missing values
• 2. Fill nulls for 'LotFrontage' with median value calculated after grouping by 'Neighborhood'
• 3. Fill nulls for 'GarageYrBlt','MasVnrArea' with 0
• 4. Apply log-transform on target feature 'SalePrice'
• 5. One-hot encoding

3 Train Linear Regression

Split dataset in training set (X_train, y_train) and test set (X_test, y_test)

4 Evaluate Linear Regression model

R^2 score on trainig set: 0.94609, MSE score on trainig set: 0.00808

R^2 score on test set: 0.89136, MSE score on test set: 0.01472

5 Model refinement with Ridge regression and Lasso regression

Ridge regression (alpha=0.05): R^2 score on training set: 0.94598, R^2 score on test set: 0.89410

Lasso regression (alpha= 0.0001): R^2 score on trainig set: 0.94169, R^2 score on test set: 0.90843

6 Conslusion:

6.1 In practice, ridge regression is usually the first choice between two models.

6.2 However, if you have a large amount of features and expect only a few of them to be important, Lasso might be a better choice.