Suppose we have a set of labeled and unlabeled instances. 70%unlabeled 30% labeled. We apply a semi-supervised algorithm. Let's say we apply S3VM or Laplacian SVM. We use all the data available. When finish training we have a function $f(\mathbf{x})$ that predict the class label. Is it correct now to label the unlabeled samples we have used to train the classifier with this function?
1 Answer
$\begingroup$
$\endgroup$
2
First: S3VM and LapSVM are NOT label propagation algos.
Second: it is never this simple. For the S3VM , you usually need to know the fraction of (+) labels in the unlabelled set. It is certainly true for the TSVM in SvmLin. I don't know if LapSVM will be skewed if the Unlabelled set has a very different balance.
I have a blog post on this
-
$\begingroup$ Thanks, it helped me the post. I'm going to update my first post. $\endgroup$– user45299Commented Nov 24, 2014 at 7:04
-
$\begingroup$ How can I apply feature selection when using TSVM or manifold regularization $\endgroup$– user45299Commented Nov 24, 2014 at 7:13