Is building a multiclass classifier better than several binary ones? I need to classify URLs into categories. Say I have 15 categories that I'm planning to zero down every URL to. 
Is a 15-way classifier better? Where I have 15 labels and generate features for each data point. 
Or building 15 binary classifiers, say: Movie or Non-Movie, and use the numbers I get from these classifications to build a ranker, to pick the best category, going to be better? 
 A: This will depend on how your data is dispersed.  There is a beautiful example that was given recently to a similar question where the OP wanted to know if a single linear discriminant function would be a better classifier for deciding population A vs B or C or one based on mutliple linear discriminant functions that separate A ,B and C.  Some one gave a very nice colored scatterplot to show how using two discriminants would be better than one in that case.  I will try to link to it.
A: First of all, you must ask yourself if your problem is multilabel (i.e. a single URL can belong to several classes) or not (i.e. a single URL can belong to only one class). 
If the former, go with a battery of binary classifiers, because this is a default way of doing  multilabel problems.
If the latter, the answer depends on a combination of how does your data look, what is the aim of your analysis and what method are you using -- probably you should just try both and select best.
Only note that some methods (like SVM) can't actually do multiclass classification because of how they are defined and thus internally use a battery of binary classifiers.
A: Some methods deal well with multiclass, Random Forests, MLPs for example.
If you don't want to go that way, then it is possible that ECOC may well out perform 1-vs-All for your problem, only testing will tell.
