On UCI adult is 4mb, on kaggle it's 400k?
https://archive.ics.uci.edu/ml/machine-learning-databases/adult/
https://www.kaggle.com/uciml/adult-census-income

(kaggle is zipped)

I'm not sure about missingness patterns there

You want a dataset with missing values completely at random ? Patterns in missing values could be interesting too.

I guess I would be most interested in not missing at random. The main point was to have a non-trivial amount of missing data and possibly across multiple features.
Any dataset that's both small enough and interesting enough should be fine, I wouldn't worry about trying to do statistics (which are probably impossible to compute given the data is actually missing).

I just realized the biggest problem with this issue: we can either return a weirdly typed numpy array or loading the dataset will have a soft requirement on having pandas installed....

@jnothman Sorry I don't think I understood your statement.

Yes, my statement was not a complete sentence.

I was basically saying I don't think we especially need a real-world dataset with missing values at this point.

I think we should make it possible to fetch Titanic as a dataframe (#11875).

And I think we should make it possible to insert missing values in a not-completely-at-random fashion (#6284).

I agree, we should be able to fetch titanic. I was a bit disappointed when I realized we're still doing a pandas read_csv to my github repo in an example.

Maybe the ames housing data would be appropriate, not sure if it has missing values.

The Ames housing dataset has missing values in the CSV files that can be found on Kaggle (login required to download the data):

https://www.kaggle.com/c/house-prices-advanced-regression-techniques/data

There is also a version on OpenML but it does not have missing values:

https://www.openml.org/d/41211 It seems that they have been imputed by zeros.

The openml version is apparently derived from this R package:

https://cran.r-project.org/web/packages/AmesHousing/index.html

but the data is in rda format. I am not sure if the missing values are there.

If you are interested I trained some baseline models + model inspection in the following notebook:

https://nbviewer.jupyter.org/github/ogrisel/notebooks/blob/master/sklearn_demos/ames_housing.ipynb

BTW the original dataset is there:

• http://www.amstat.org/publications/jse/v19n3/decock/AmesHousing.xls
• http://www.amstat.org/publications/jse/v19n3/decock/AmesHousing.txt.
• http://www.amstat.org/publications/jse/v19n3/decock/DataDocumentation.txt

and the paper describing the motivation, design and analysis of this dataset: http://jse.amstat.org/v19n3/decock.pdf

OpenML allows having several versions of a dataset, so we could push a version with the missing data.
For the Ames Housing dataset, you need to log and remove outliers. If you don't remove outliers, your cross-validation score will have crazy variance in my experience.

Actually as @maikia found in #16345 there is already another version of the same dataset. But it does not have the missing values either. Maybe the ARFF format is limiting this? I haven't investigated.

I did not observe crazy variance in the notebook above but maybe the outliers have already been removed?

In your initial prototype they are in the training set, and I guess gradient boosting doesn't care as much.
ARFF definitely supports missing values.
These guys here:

See here:
https://github.com/amueller/COMS4995-s20/blob/master/slides/aml-05-linear-models-regression/aml-05-linear-models-regression-Copy1.ipynb

If you do a linear model, I would be curious to see how performance varies across cross-validation splits.
This is what I got for ridge:

Actually as @maikia found in #16345 there is already another version of the same dataset. But it does not have the missing values either. Maybe the ARFF format is limiting this? I haven't investigated.

I think there are quite few missing values in this dataset
https://www.openml.org/d/42165