Regression datasets from the UCI machine learning repository prepared for benchmarking studies with test-train splits.

Setup

This repository uses git large file storage so you must first install git LFS otherwise the cloned repo will only contain pointer files rather than the data files. The size of the repository is about 319 Mb.

After installing git LFS you can simply clone and setup the python package as follows:

git clone https://github.com/treforevans/uci_datasets.git
cd uci_datasets
python setup.py develop

Note that you must use develop in the above line, not install.

Usage

The following code gets the first test-train split (i.e., split=0) of the challenger dataset:

from uci_datasets import Dataset
data = Dataset("challenger")
x_train, y_train, x_test, y_test = data.get_split(split=0)

There are 10 test-train splits for each dataset (as in 10-fold cross validation) with 90% of the dataset being training points and 10% being testing points in each split. The split parameter of the Dataset.get_split method accepts integers from 0 to 9 (inclusive).

Datasets

The below table contains the size (number of observations) and the number of input dimensions of each dataset.

Dataset name	Number of observations	Input dimension
3droad	434874	3
autompg	392	7
bike	17379	17
challenger	23	4
concreteslump	103	7
energy	768	8
forest	517	12
houseelectric	2049280	11
keggdirected	48827	20
kin40k	40000	8
parkinsons	5875	20
pol	15000	26
pumadyn32nm	8192	32
slice	53500	385
solar	1066	10
stock	536	11
yacht	308	6
airfoil	1503	5
autos	159	25
breastcancer	194	33
buzz	583250	77
concrete	1030	8
elevators	16599	18
fertility	100	9
gas	2565	128
housing	506	13
keggundirected	63608	27
machine	209	7
pendulum	630	9
protein	45730	9
servo	167	4
skillcraft	3338	19
sml	4137	26
song	515345	90
tamielectric	45781	3
wine	1599	11

Dataset information can be obtained from the all_datasets dictionary. For example, to obtain a list of all datasets with fewer than 1000 observations, execute the following:

from uci_datasets import all_datasets
[name for name, (n_observations, n_dimensions) in all_datasets.items() if n_observations < 1000]

Papers using these datasets

The following papers use the same datasets and test-train splits present in this repository.

• Yang, Zichao, et al. "A la carte–learning fast kernels." Artificial Intelligence and Statistics. 2015.
• Wilson, Andrew Gordon, et al. "Deep kernel learning." Artificial Intelligence and Statistics. 2016.
• Evans, Trefor W., and Prasanth B. Nair. "Scalable Gaussian processes with grid-structured eigenfunctions (GP-GRIEF)." International Conference on Machine Learning. 2018.
• Evans, Trefor W., and Prasanth B. Nair. "Discretely relaxing continuous variables for tractable variational inference." Neural Information Processing Systems. 2018.
• Audouze, Christophe, and Prasanth B. Nair. "Sparse low-rank separated representation models for learning from data." Proceedings of the Royal Society A 475.2221 (2019).
• Evans, Trefor W., and Prasanth B. Nair. "Quadruply stochastic gaussian processes." arXiv preprint arXiv:2006.03015 (2020).