I'm working on classification of delay / (no delay) / negative delay for airplanes ( 2 or 3 labels). I'm also playing with idea to model no delay, delay and large delay.

I noticed that there are cases where I can quite well train neural network form scikit that it makes little mistake for either delay or negative delay in the training set. From my little experience my gut is telling me it is nonsense to use two/three models for each label. What if they disagree on the label? Could I then pick the one with highest probability or something?

I dont think it is a good approach, but cannot find any articles explaining it.

I'd like to check, is it non-sense to have multiple models in multi-label case or not? If yes or no, why?

Many thanks!


What if they disagree on the label?

There are other ways of combining classifiers. For example, it's possible your labels display a natural hierarchy. A flight might either be delayed or not. If it is delayed, it might have a short delay or a long delay (which is an example you give); it it is not delayed, it might be on time or before time (another example you give).

Breaking down your multilabel classification into a hierarchy of problems might make sense in some case. The dataset sizes of a node's children might be very skewed. The misclassification costs might be hierarchical in nature. Along the way, this bypasses your problem of classifier disagreement.

See, for example Multiple Model Approaches to Modelling and Control, by Roderick Murray-Smith and Tor Arne Johansen, and Combining Artificial Neural Nets, by Amanda J. C. Sharkey. This is sometimes referred to as modular classifier combination.

  • $\begingroup$ Thank you very much, I will have a look at the papers. So I take it, it is possible to have nested models like this, they just need to be hieararchy kind of type, right? $\endgroup$
    – Jan Sila
    Oct 13 '16 at 7:38
  • $\begingroup$ @JanSila Yes. Of course, these approaches work better when the hierarchy is natural and inherent to the problem (which seems plausible in your case). $\endgroup$
    – Ami Tavory
    Oct 13 '16 at 8:25
  • $\begingroup$ @JanSila You're welcome. All the best. $\endgroup$
    – Ami Tavory
    Oct 13 '16 at 10:02

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.