1
$\begingroup$

I need a better classifier for an unbalanced multi-class problem. I'm tempted to implement Adaboost as an exercise in understanding it since it is quite clear and simple, but am unsure how to deal with unbalanced samples in this context. Can anyone recommend any good introductions and/or papers that deal with unbalanced datasets in a multi-class boosting context? Alternatively, can anyone suggest a suitable free (online?) classifier that I could submit my data to to get a sense of how well an Adaboost might work on it?

$\endgroup$
1
  • $\begingroup$ From my understanding of your post, you are asking three questions, and adressing all of them would require too long of a post. It would help to narrow the focus of your question to attract more direct answers. Candidate subquestions could be "How to deal with multiclass classification in Adaboost" (Adaboost itself is not multiclass, framing the question for the general gradient boosting would help), "How to deal with unbalanced classes", "What boosting libraries are available". If you search for each of these question individually, you are likely to find relevant answers here or on Google. $\endgroup$
    – Winks
    Jul 2, 2016 at 8:07

1 Answer 1

1
$\begingroup$

Since this was as much an exercise in understanding Adaboost as anything, I chose to split the problem into several single-class classification tasks, and use a nearest neighbour algorithm to identify the "closest" N negative samples to contrast with the N positive samples for each class. While the results were not spectacular, it gave me the insight into the workings of Adaboost that I was after, and also, courtesy of the nearest neighbour pre-processing, some insights into how my data was distributed. Naturally, I would hope this is not the best answer there is...

$\endgroup$

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.