https://archive.ics.uci.edu/ml/datasets/Liver+Disorders lists some papers which have used this dataset. I went to one, Turney, and it says:

The target concept was defined using the sixth column: Class 0 was defined as “drinks < 3” and class 1 was defined as “drinks ≥ 3” Column seven was originally used to split the data into training and testing sets.

In support of this, the file bupa.names says

  1. selector field used to split data into two sets.

This doesn't sound like a dependent variable.

But I went to another, Jiang and Zhou and it says there are 6 attributes (hence, presumably, the seventh is the one they use as the dependent variable).

I also found this which claims that the seventh variable is:

a dependent variable with two levels, indicating presence or absence of a liver disorder

Are a lot of people doing classification on a completely random variable?


Some time later, I can now answer my own question. I contacted the person who is named as donating the original data, Richard Forsyth. He confirmed that the final variable is not a dependent variable, but is intended as a train/test split. We wrote a short letter to Pattern Recognition Letters to point this problem out. A few more details here: http://jmmcd.net/2016/04/30/disorder.html

In summary, A LOT of people are doing classification on a variable that they think indicates liver disorder, but is actually just a train/test indicator. When I say A LOT, I mean hundreds, including in top ML venues. Also some people are drawing conclusions about liver disorder diagnosis and treatment.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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