Suppose there are bags of unlabeled data and I try to learn the positive instances from those bags. I know there are lots of bag based classification algorithms but my region of interest is positive instance selection based methods like MI-SVM.
APR, Diverse Density, EM-DD, miSVM, MILBoost can all give you instance labels (either through a density function, or directly). MILES is a bag-based method I suppose, but you can also get instance labels with it: it selects discriminative instances, which you can assume are the concept instances.
Maybe this paper is also interesting: http://machinelearning.wustl.edu/mlpapers/paper_files/AISTATS07_GehlerC.pdf
What kind of application of MIL are you considering?
Given the unlabelled data, I try to uncover positive instances as defining the classifier.
I know one of the application of Deterministic Annealing to Multiple instance learning over Random Forests.
What I mean is, what are your bags and what are your instances?
- Badges
- Science topic