Skip to content

aejsong/psv-ameshousing

Repository files navigation

Ames Housing Data and Kaggle Challenge

Overview

The objective of this project is to create a Linear Regression Model that predicts the sale price of properties in Ames, Iowa. A model is created with training data that was provided to us and predictions will be made on a test data that does NOT have the actual sale prices available to us. The predictions will be submitted onto Kaggle and our Root Mean Squared Error will be calculated for our predictions vs. actual values.

Kaggle Competition

Predicted sale prices for test data without sale prices were uploaded onto a Kaggle competition where Root Mean Squared Error for those predictions were returned.
https://www.kaggle.com/c/dsi-us-9-project-2-regression-challenge/leaderboard#score

Data Dictionary

Having this link open while going through the code may be helpful in understanding the specific features: http://jse.amstat.org/v19n3/decock/DataDocumentation.txt

Data Cleaning

Many of the columns had NaN values in them.

  • Rows that had a NaN value as a string, I cross checked if the area for that specific variable was 0 sq. ft. and concluded that the NaN values could be converted to NA. (ie. NA = No basement 1, No pool, etc.)

Some rows had an entire categorical variable missing. (ie. Basement, Basement Sq. Ft., Basement Full Bath, etc.)

  • I concluded that these variables are important to help determine the price of a property and decided to remove rows like these so they wouldn't affect my model.

Outliers:

  • While looking at my outliers post-modeling, I noticed that some properties were huge with many rooms and bathrooms but were sold at an extremely low price. I decided to drop these outliers.
  • There were other outliers that were identified as outliers in multiple categories. These outliers were also dropped.

Model Process

  1. I dummied out every single categorical column in both train and test data.

  2. I split my training data into one part that I could use to train my model and the other party so I could test it. In order to not clash with my original test data, I labeled this new test data as "holdout."

  3. Identified variables I want to include my model.

  • I first started by determining which variables are definitely important to determining sale price of a property.
  • I created a for-loop that made Single Linear Regressions for each individual variable against Sale Price and added that variable name to a list if it affects sale price by at least +/- 30,000.
  1. I ran my model and obtained the scores of the training set, the holdout set, and a cross value score of the training set.

  2. I used PolyNomialFeatures to create interactions of every numerical column in my original dataset. I used Lasso to identify which interactions were the most effective and I added those to my Linear Regression Model and re-ran it.

Summary

Multiple submissions were made on Kaggle. My best model has a Root Mean Squared Error of 23,808.780 against the Kaggle test data.

The following are scores obtained from my training and holdout data:

Metric Score
R2 Score (Train Data) 93.95%
R2 Score (Holdout Data) 92.11%
Cross Value Score, 5 Folds (Train Data) 92.00%
Root Mean Squared Error $21,941.01
Mean Absolute Error $15314.41
Mean of Residuals $758.53

Conclusion

For the most part, I believe my model does ok. Based on the metrics, it doesn't seem like it's overfitting too much. For future note, I will put in more time to separating/creating bins for the data by methods such as numerical vs. categorical or grouping properties into a different column in order to reduce the amount of dummy columns I have. I feel if I had done that, my model would be a little bit better.

Most of the properties on the training set were in the range of $100,000 to $400,000. If I felt that my property should be priced within that range, I would be fairly confident in using my model to predict it's value. However, due to the overwhelming amount of average priced properties, this makes the model less accurate for properties that are worth more. If my property is greater than $400,000 I wouldn't use my model to sell my property and go to a real estate agent instead.

Sources