locked
Choosing Classification Label and order in which it is processed (Migrated from community.research.microsoft.com) RRS feed

  • Question

  • Hari1234 posted on 02-13-2011 9:55 PM

    So I was playing with Multi Point Bayes Classifier and noticed that changing the classifier label for my test seem to have some effect (though I would not say very large but does seem to have considerable effect on predictions/probabilities). So I have 4 classifications in my training data and I have labelled them 0,1,2,3. When debugging I wanted to look at the weight vectors and found that for class "0" it is always 0..... (it probably has to do with following code as it checks for c==0)

    for ( int c = 0; c < nClass; c++)
    {
        trainModel.wInit[c].ObservedValue = (c == 0)
        ? VectorGaussian.PointMass(Vector.Zero(nFeatures))
        : VectorGaussian.FromMeanAndPrecision(Vector.Zero(nFeatures)
    }
     

    So I wanted to find out what would happen if I relabelled all my test data in such a fashion that previous "0"'s are now "3" and previous "1"s are now "2" ..and so on.. While the results seem to be some what consistent for majority part with my test data, however there are some differences in probabilities. This I am thinking probably because the regions are intersecting at slightly diffferent points and  due to the above initialization of point mass.

    This leads to how should one define class labels and what factors needs to be taken into account because renaming the labels slightly differently does seem to have an effect on probabilites (predictions)

     

     

     

    Friday, June 3, 2011 6:18 PM

Answers

  • John Guiver replied on 02-18-2011 10:37 AM

    Yes you are correct. The reason this line is in the code is to remove a redundant degree of freedom, and this does introduce asymmetry.

    There is likely to be some way to remove the degree of freedom in a symmetric way. For example making the weights sum to zero may be a better way to handle this, but we have not investigated this yet. Thanks for bringing this to our attention

    John G

    Friday, June 3, 2011 6:18 PM