mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <>
Subject Re: multi-class classification question
Date Sat, 03 Aug 2013 06:59:37 GMT
On Fri, Aug 2, 2013 at 6:28 AM, yikes aroni <> wrote:

> Hmm. Interesting indeed. And the real bonus is that i get it. Sounds like
> the solution to the problem of large numbers of target classes is to
> develop a sensible scheme for grouping the target classes s.t. you can
> reduce the number of them required per training run, then run the training
> once on each sub-set of target classes. To query, you simply query this
> large number of classifiers to build up your answer.

You may have thousands of classifiers in toto, but you should only have to
evaluate dozens of them.

> Is that the common practice?

Having training data for 60,000 categories is quite rare in my experience.

> thank you for your time, Mr. Dunning.

NP.  Send feedback if you implement such a beast.

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message