37
$\begingroup$

What are the freely available data set for classification with more than 1000 features (or sample points if it contains curves)?

There is already a community wiki about free data sets: Locating freely available data samples

But here, it would be nice to have a more focused list that can be used more conveniently, also I propose the following rules:

  1. One post per dataset
  2. No link to set of dataset

  3. each data set must be associated with

    • a name (to figure out what it is about) and a link to the dataset (R datasets can be named with package name)

    • the number of features (let say it is p) the size of the dataset (let say it is n) and the number of labels/class (let say it is k)

    • a typical error rate from your experience (state the used algorithm in to words) or from the litterature (in this last case link the paper)

Improve this question
$\endgroup$
3
  • $\begingroup$ +1, but the ones from NIPS2003 have train.labels only -- the NIPS2003 paper says clearly "validation and test set labels are withheld". $\endgroup$
    – denis
    Oct 24, 2011 at 10:00
  • $\begingroup$ Thanks. The comment about NIPS is for the answer from @mbq . $\endgroup$
    – robin girard
    Oct 24, 2011 at 10:02
  • $\begingroup$ Anyone here have a high dimensional dataset with more than two class labels? $\endgroup$
    – hlin117
    Nov 14, 2015 at 2:48

5 Answers 5

Reset to default
3
$\begingroup$

Dorothea
n=1950
p=100000 (0.1M, half is artificially added noise)
k=2 (~10x unbalanced)
From NIPS2003.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Can you explain how this is 100000 features? I look at the training data and each line has maybe 2500 integers per line. $\endgroup$
    – JeremyKun
    Jan 19, 2016 at 0:03
  • $\begingroup$ It is a sparse array, integer N means attribute N value is 1. $\endgroup$
    – user88
    Jan 19, 2016 at 14:27
3
$\begingroup$

Gisette
n=13500
p=5000 (half is artificially added noise)
k=2 (balanced)
From NIPS2003.

Improve this answer
$\endgroup$
3
$\begingroup$

Dexter
n=2600
p=20000 (10k+53 is artificial noise)
k=2 (balanced)
From NIPS2003.

Improve this answer
$\endgroup$
3
  • $\begingroup$ I don't understand... one set per person? $\endgroup$
    – user88
    Jul 30, 2010 at 17:40
  • $\begingroup$ @robin & @mbq I would suggest keeping it one dataset per post. This so people can indicate with votes which of the suggested ones there also suggest/support $\endgroup$
    – Peter Smit
    Jul 30, 2010 at 17:59
  • $\begingroup$ @Peter, OK, I follow your idea, I have changed the question accordingly. $\endgroup$
    – robin girard
    Jul 31, 2010 at 6:14
3
$\begingroup$

Arcene
n=900
p=10000 (3k is artificially added noise)
k=2 (~balanced)
From NIPS2003.

Improve this answer
$\endgroup$
2
$\begingroup$

Prostate (gene expression array)

  • k=2
  • n=48+52
  • p=6033

Available via (among other) R package spls name of the dataset: prostate

error rate = 3/102 (see here) also I think there are paper which show 1/102 error rate. I would say this is an easy test case.

Improve this answer
$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.