0
$\begingroup$

I've been tasked with creating a model that can classify individual cases as either something to be flagged or not. However, the training data I have access to is only aggregate data, where each row is the number of cases in a particular state. I'm much more familiar with training on data with individual observations vs aggregate, so I'm a little stumped on where to begin.

Is this even statistically feasible? I've tried looking into learning from label proportions, but I'm not sure it fits exactly what I'm looking to do here.

Additionally, there's been talk of performing clustering (k-means clustering) but not sure if I'd end up with genuine clusters of similar cases or if I'd just end up with clusters of similar states.

Improve this question
$\endgroup$
3
  • $\begingroup$ Yes it's quite standard. You need to look for models that take sample weights. Your cross validation has to be altered too... $\endgroup$
    – seanv507
    Commented Jan 27, 2021 at 1:19
  • $\begingroup$ I have done this sort of thing with logistic regression and it turned out to be easier than I thought - just a matter of weighting $\endgroup$
    – Henry
    Commented Jan 27, 2021 at 1:41
  • $\begingroup$ But if you want any more concrete advice on what to do, you must give us some context and details, especially what aggretagte stats you have. $\endgroup$
    – kjetil b halvorsen ♦
    Commented Jan 27, 2021 at 20:34

0

Reset to default

Your Answer

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