Orange Forum • View topic - Multi-class Classification

Multi-class Classification

A place to ask questions about methods in Orange and how they are used and other general support.

Multi-class Classification

Postby qingshuang » Thu Aug 15, 2013 1:58

I used the Classification trees, RF (random forest), kNN (k-nearest neighbor) and NB (naïve Bayes) methods in Orange Canvas for multi-class classification problem, do they all support multi-class classification? which stratagy do they use (detailed stratagy of each individual classifier)?

Re: Multi-class Classification

Postby Ales » Mon Aug 19, 2013 13:44

qingshuang wrote:do they all support multi-class classification?

Yes

qingshuang wrote:which stratagy do they use

All of the mentioned methods support multi-class classification intrinsically. Please refer to the documentation and relevant references.

Re: Multi-class Classification

Postby qingshuang » Mon Aug 26, 2013 8:49

Ales wrote:All of the mentioned methods support multi-class classification intrinsically. Please refer to the documentation and relevant references.

I also believe they support multi-class classification intrinsically, but I have a question about kNN multi-class classification. In binary classification, an object is classified by a majority vote of its neighbors, in multi-class classification, sometimes, the vote could be equal. For example, an object, the 5 nearest neighbors, 1 of the neighbors belongs to class"1", 2 belongs to class "2", and 2 belongs to class "3", then which class will the object classified?

Re: Multi-class Classification

Postby Ales » Mon Aug 26, 2013 9:59

A random choice between class "2" and "3" (but a consistent random choice i.e. the instance for which the classification is sought is used to seed the random generator, so classification on the same instance always returns the same class).

Re: Multi-class Classification

Postby qingshuang » Mon Aug 26, 2013 12:09

Ales wrote:A random choice between class "2" and "3" (but a consistent random choice i.e. the instance for which the classification is sought is used to seed the random generator, so classification on the same instance always returns the same class).

Will the other classifier(i.e. RF, NB and CT) have the same problem (cannot give a exact choice but a random choice)?

Re: Multi-class Classification

Postby Ales » Fri Aug 30, 2013 10:56

qingshuang wrote:Will the other classifier(i.e. RF, NB and CT) have the same problem (cannot give a exact choice but a random choice)?
Naive Bayes and Classification trees use the same strategy, but Random forest seems to break ties by always picking the first candidate (this should probably be fixed and made consistent with others).


Return to Questions & Support



cron