0
$\begingroup$

I'm trying to model a binary classification problem. 5 continuous features, slightly imbalanced dataset (about 60 in one class, a little more than 200 in another).

So far I've tried kNN, LDA, and C5.0. C5.0 works very well (e.g. the AUC for ROC is very good).

However, I've read that boosting a la AdaBoost has problems in the presence of mislabeled data, and that C5.0's boosting is reminiscent of AdaBoost. My data is mislabeled insofar as perhaps 5% of them are probably wrong. (I'm vague here, because it's a real-world problem where the ground truth is fuzzy---good, but not perfect.)

Does anyone know how well C5.0 boosting performs in the presence of mislabeled data?

$\endgroup$
  • $\begingroup$ The problem is that nobody has really studied C5.0 so far (it has been in the private domain until a year or so ago). $\endgroup$ – topepo Nov 23 '13 at 5:06
2
$\begingroup$

AdaBoost is indeed susceptible to mislabeled data (aka noise). Friedman et al's paper "Additive logistic regression: a statistical view of boosting" offers an explanation.

However, C5.0's boosting is somewhat different from AdaBoost, e.g. additive rather than multiplicative weight adjustment of misclassified cases, with the goal of improving performance in the presence of moderate noise.

C5.0 was released under GPL in Jan 2011.

$\endgroup$

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.