3
$\begingroup$

Is there a term for the inaccuracy that results from an ML model being trained on imperfectly labelled data? For example, if humans label a training set, they could make occasional human errors. In such cases, a theoretically perfect model trained on that dataset would still have some inaccuracy (purely due to that human error during labelling). Note that human error may not be the only cause of inaccurate labels; it's just a simple example.

Example use of the term: "the data set was labelled by humans, so we should anticipate some observations being incorrectly labelled, and hence a model that performs with some degree of error"

$\endgroup$
2
  • $\begingroup$ Also, maybe related: datascience.stackexchange.com/questions/17839/… $\endgroup$
    – stevec
    Commented May 11, 2020 at 8:55
  • 1
    $\begingroup$ This is sometimes called "label noise," the acknowledgement that the labels themselves may be incorrect (for any reason, including human error). $\endgroup$
    – Sycorax
    Commented Jul 21, 2020 at 15:43

1 Answer 1

2
$\begingroup$

Answered in comments by Sycorax:

This is sometimes called "label noise," the acknowledgement that the labels themselves may be incorrect (for any reason, including human error).

$\endgroup$

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.