1

I'm currently working on a paper that I want to evaluate on a publically available dataset. The following requirements apply:

  • Classification (target variable is boolean or a factor)
  • n observations of m objects
  • 1:m of the objects are observed multiple times (the more, the better)
  • there exist objects m that are always classified as 0 and objects that are observed at different states (0/1)
  • every object m has a unique id (statements like "object a had class 0 at observation n_1 and class 1 at observation n_2" are possible)

E.g.

+------+----------+----------+----+
|class | feature1 | feature2 | id |
+------+----------+----------+----+
|    1 |      1.1 |      0.3 |  a |
|    0 |      0.8 |      0.4 |  a |
|    0 |      0.9 |      0.3 |  b |
|    1 |      1.0 |      0.3 |  c |
+------+----------+----------+----+

Does anybody know a dataset that matches the criteria and is able to share a (link to a) .CSV? E.g. to the UCI repository.

An example is the forrest fire data set. This dataset holds several "forrests" (m, identified by coordinates) and the target is to predict if there was a fire or not. Almost Every forrest m is observed more than once - which is perfect.

Improve this question

2 Answers 2

Reset to default
3

Not sure what you're trying to do so this could be totally out there, but what about finance data like stocks?

  • Objects m are unique company stock symbols like GOOG
  • Classifications are the stock price trend at every sequential data point, positive or negative, represented as a boolean 1 or 0.
  • Observations n are the trends at discrete points in time
  • Not sure what your features are suppose to be, but price could be one

You can have as many m and as many n for each m that you want. Data sets are here, as well as a ton of other non-finance related ones if you don't like that idea.

Improve this answer
4
  • It's a bit unclear what features to use. I don't want to spend to much time in feature engineering since I want to use the dataset for a scientific paper (where I would have to explain the whole process and that's too much wasted space)
    – Boern
    Commented Apr 7, 2017 at 15:05
  • @Boern Is there a reason you can't use the forest fire data you just linked to? They have a CSV file for the data. Also it'd be helpful if you explained what you mean by "features" or what kinds of things you're looking for. To me that just means interesting observations in some number value range.
    – timcardenuto
    Commented Apr 8, 2017 at 21:49
  • My backup plan is to use the forrest dataset, although not being ideal because alsmost every object/forrest (except 2/3, 8/4 and 9/5) is observed multiple times. A dataset with a higher ratio of objects, observed only once, is desirable. We have the same understanding of what "features" are. I just don't get what to use as feature in your proposed dataset. The "price" is too trivial. Also, more features would be desirable.
    – Boern
    Commented Apr 10, 2017 at 7:27
  • Another caveat of the forrest dataset is, that is it is intended to be used for regression (instead of classification)
    – Boern
    Commented Apr 10, 2017 at 14:01
2
+25

How about bike share data?

https://archive.ics.uci.edu/ml/datasets/Bike+Sharing+Dataset (currently the UCI website is down for me - 4/9/17)

https://www.kaggle.com/c/bike-sharing-demand/data

Each station (m) has multiple observations about how many bikes are checked out, and it has features like temperature, humidity, season, whether it's a holiday, etc.

Improve this answer
2
  • It's hard to tell if the dataset fits (since UCI is down). According to kaggle, there is no "station ID", allowing me to identify the the different stations (m). Please prove me wrong.
    – Boern
    Commented Apr 10, 2017 at 7:17
  • How about the Bay Area bike share data? This has station and bike IDs. bayareabikeshare.com/open-data
    – szxk
    Commented Apr 10, 2017 at 15:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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