1
$\begingroup$

As stated in the title I would like to find some classification datasets with $c \ge 5$ classes and one categorical feature or more. It is even fine if there are some continuous features but I would like at least a categorical one.

I have some genetic datasets where features take values in $\{ G, T, A, C\}$ but unfortunately the class is binary.

I have a method that should improve the accuracy performance on those dataset. I used some of UCI datasets but I need more of them to show a significant $p$-value that states the improvement of performances.

Improve this question
$\endgroup$
3
  • $\begingroup$ Let me see if I get this: you get a performance increase for every data set. And then you use some test statistic that aggregates those performance increases? And now you need more data sets so you can get a higher sample size for calculating the test statistic und thus get a low p value for the test statistic? $\endgroup$
    – Momo
    Oct 8, 2012 at 1:00
  • $\begingroup$ Yes, I would like to have more datasets too see if I can a low p-value for the test statistic. $\endgroup$
    – Simone
    Oct 8, 2012 at 1:02
  • 1
    $\begingroup$ OK, thanks. As a side note: Unless I am missing something, I don't think that is a valid approach. You need to have the sample size fixed before you conduct your test, else you invalidate the statistical analysis imo, as you now found the p value was to high and need new data. However, you could think of using a sequential test procedure that adds the result from one additional dataset, than calculates the test than tells you H0, Ha or look for more data. $\endgroup$
    – Momo
    Oct 8, 2012 at 1:11

1 Answer 1

Reset to default
2
$\begingroup$

Check out the UCI ML repository http://archive.ics.uci.edu/ml/

It has more than 150 data sets for various classification tasks and serves as a well accepted collection of datasets for benchmarkng new methods. I'm sure you'll find a multiclass response data set with categorical predictors.

Edit: I'm sure you know, but let me just mention that if you have predictors with 5 or more classes nothing prohibits to take a predictor as the target.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks. I've already retrieved them all. But I don't have enough datasets to sketch a statistical test. $\endgroup$
    – Simone
    Oct 7, 2012 at 23:31
  • $\begingroup$ Alright, but you didn't find any there? Also what do you mean with "not enough datasets to sketch a statistical test"? More info on what you need/want would probably help us help you better. $\endgroup$
    – Momo
    Oct 7, 2012 at 23:34

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