Hi all,
I have the following problem (and I hope that I will describe it understandably):
I am processing a dataset containing polynomial labels including n-1 "positive" labeltypes and one negative (standard) label-type.
I am using the Binary2MultiClassLearner because the learner I use is a binary SVM.
The problem now exactly is, that the "negative" class dominates the dataset and the SVM just predicts the "negative" class .
For binominal labels it is possible to define the positive class. It would be nice if one could define the negative class for my problem, is this possible?
Thank you.
Best,
schalkekid
As you surmised, SVMs default to the most frequent label value when no separating pattern is found. You can try the EqualLabelWeighting operator, which beefs up the importance of examples with rare labels, or you can construct a new label attribute in which all the non-negative values are merged, and then bin the old label attribute. In the latter case you would use the AttributeConstruction operator, which is discussed below in this forum...
[url=http://rapid-i.com/rapidforum/index.php/topic,580.0.html[/url], click to go there.