Suppose I have skewed distributions of classes in train set.

How should I deal with it? Just train and network will deal itself? Or some methods are good?

For example, can I artificially make distribution uniform, then train my network and then apply some bayessian methods to take probabilities into account?


Distributions of classes is following:

enter image description here

This is from Planet: Understanding the Amazon from Space competition on Kaggle. Plot is from this kernel. Multilayered convolutional network is used.

  • $\begingroup$ By skewed distributions did you mean class imbalance? I'm weak with terminology. $\endgroup$ – papabiceps Jul 5 '17 at 4:20
  • $\begingroup$ @papabiceps yes, sorry, I am weak in terms too; some classes are very often and anoter ones are very rare... $\endgroup$ – Dims Jul 5 '17 at 9:03
  • $\begingroup$ What kind of neural network are planning to use and what kind of data are you trying to classify?Do you have any training algorithm in mind ? And how skewed is your dataset can you tell us the proportions like 95% of positive class and 5% negative class. $\endgroup$ – papabiceps Jul 5 '17 at 10:44
  • $\begingroup$ @papabiceps see my update please $\endgroup$ – Dims Jul 5 '17 at 16:40

Your Answer

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

Browse other questions tagged or ask your own question.