1
$\begingroup$

I have a dataset of 1000 labels. I would like to build a custom model from this dataset with three labels, say Dog, Cat, and Others. Obviously, Others will be large in number for this approach. I understand that this could be an issue since the classifier has seen more Others examples.

Pipeline

Large dataset --> Seperate to Dogs, Cats and others --> Classifier --> Predicts Dogs, Cats or Others.

What are the other approaches which can be taken in this situation?

$\endgroup$
1
  • 1
    $\begingroup$ More importantly it is to know the distribution of labels, how many of dog, cat, etc. are there? I would say if the approach you explained using 'Others' makes the labels highly imbalance, maybe look at the frequency of top N labels by which the distribution of labels still remains balanced using quartile (e.g. 0.25, 0.5, 0.75)! $\endgroup$ Aug 9, 2018 at 15:34

1 Answer 1

1
$\begingroup$

One option is to change the label distributions. If you want Others to be proportional to Dog and Cat, you can randomly drop instances that are not Dog and Cat.

Another option is to pick a classifier that is not influenced by the number of instances, such as support vector machine (SVM).

Yet another option is to use evaluation metrics to examine performance at the individual category level (e.g., precision for Dog, Cat, Others separately).

$\endgroup$

Your Answer

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

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