I am working on a classification problem with 7 classes. Is there any rationale to suspect that the best model might be found with a multiclass classifier, multiple one-vs-all classifiers, or even a set of one-vs-one classifiers?

The 7 one-vs-all classifiers seems conceptually the simplest to describe each class individually. However, if each class is highly distinct, I suspect that the classifier might have trouble grouping the "rest" into a cohesive set that can be easily distinguished. In this case, I think the multiclass classifier may work best to describe each class simultaneously as being distinct from all the others. Finally, one-vs-one classification seems like it can overcome some sampling issues at the cost of computational complexity. However, this approach seems like it could be confounded by pairs of similar classes. Low-power classifiers built to distinguish these similar classes don't provide much information, but are still given the same weight as the classifiers that properly distinguish between distinct classes.

Of course, there won't be one algorithm that works best in all cases, but are there any properties of the data I might be able to use to guide me toward a good classification scheme?



Your Answer

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

Browse other questions tagged or ask your own question.