Kaggle Housing Regression Check

Based on original housing data for Ames Iowa 2011

One of Kaggle's long running "Knowledge" competitions is its House Prices competition for "Advanced Regression Techniques". The idea behind that competition is pretty simple, even if the competition itself is not. Kaggle provides two data sets:

The data sets are very similar. They both contain about 1460 rows, each corresponding to the sale of a house in Ames Iowa over the five year period from 2006 to 2010. Each row contains about 80 variables, including

One key difference between the datasets is that the training dataset includes the sale price for each house, while the testing dataset does not. To participate in the competition, you develop an algorithm that predicts the selling prices of the houses in the test set based on the other variables. You upload those predictions to Kaggle and they compare those predictions to the actual values - which they, of course, have. They then score your submission based on the root mean squared error between the logarithm of the predicted value and the logarithm of the observed sales price. That is, if there are data points and the true sale price of the data point is , while your predicted price is , then your score is Kaggle states that taking logs means that errors in predicting expensive houses and cheap houses will affect the result equally. We'll be able to see why that might make sense when we perform a regression.

The idea behind this webpage is to emulate that competition, thus allowing folks to try their hand at basic machine learning, without the need to register with Kaggle or anyone else. I've set it up primarily for my own students but I suppose anyone could try it.

This version

To try to predict housing prices here, start by downloading the train and test data sets:

These files are very close to Kaggle's train and test files but they are not exactly the same. Kaggle's data is based on an article by Dean De Cock that appeared Volume 19 of the Journal of Statistics Education. The paper and data are freely available:

I created the data files here by matching rows from the original source data to Kaggle's data. Like Kaggle, I was able to grab the actual selling prices for the test data and create a tool for you to score your predictions, even though the prices have been removed from the test data.

The data files here are not exactly the same as Kaggle's; it's really not my desire or intention for them to be the same. They are close enough, though, that your score here should give you some indication of what you might score in Kaggle's competition. The column names here are exactly the same as those in Kaggle's data - though, in a different order. That should make it easy to port your code over to Kaggle's competition, if you want. You can also read the data documentation here:

Some starter code

Kaggle provides plenty of tutorials to get you started on all of their data science competitions. I've written my own intro for beginning stats students with at least a knowledge of linear regression. This is available as the following Colab notebook:

Scoring it

To score your predictions, you should generate a predictions file in CSV format whose form looks like so:

Id SalePrice
1461 123456
1462 123457
⋮ ⋮
2916 654321

The starter notebook above generates files in exactly this format.

Once you've generated your prediction file, you can upload it using the file browser below. Your score should automagically appear.

Your score will be a non-negative number; it should be positive, since your score will be zero only if you've predicted every sales price exactly correctly. Note that smaller is better.

The histogram illustrates the scores on Kaggle's leaderboard for the competition as of the evening of May 6, 2024. Thus, you can use it to get a rough idea where your skill stand. There are a few things to keep in mind, though: