search

Dogggg / accord

Automatically exported from code.google.com/p/accord
0 stars 0 forks source link

MulticlassComputeMethod.Elimination might not produce class probabilities summing up to one. #72

Closed GoogleCodeExporter closed 8 years ago

GoogleCodeExporter commented 8 years ago 
What steps will reproduce the problem?
1. Create a multiclass machine;
2. Learn using SequentialMinimalOptimization
3. Calibrate it using ProbabilisticOutputLearning
4. Test a sample using MulticlassComputeMethod.Elimination.

What is the expected output? What do you see instead?
The probabilities (class responsibilities) may not sum up to one.

Note: Using MulticlassComputeMethod.Voting can be used as a workaround in the 
meantime.

Original issue reported on code.google.com by cesarso...@gmail.com on 19 Sep 2013 at 9:04

GoogleCodeExporter commented 8 years ago 
Well, it turns out this behavior might not be always desired. If the class 
responsibilities do not sum to one, this might be an indication that the sample 
could have been rejected. It is always possible to convert the responsibilities 
into probabilities by normalizing the responses vector to sum to one later.

Original comment by cesarso...@gmail.com on 4 Oct 2013 at 12:46