Timeline for How does a classifier handle unseen documents that do not belong to any of the pre-existing classes?
Current License: CC BY-SA 3.0
7 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 13, 2011 at 17:02 | vote | accept | Andy | ||
| May 25, 2011 at 2:54 | comment | added | Andy | That seems reasonable. Let me give it a shot. | |
| May 23, 2011 at 10:11 | comment | added | Dikran Marsupial | @andy, yes that is correct, fit a one-class classifier to the whole of the training set, and reject any observation that falls outside the boundary as being a novelty belonging to a new class. Use a conventional multi-class classifier to classify any pattern that falls within the boundary of the one-class classifier. | |
| May 21, 2011 at 22:36 | comment | added | Andy | I suppose I am a little confused now. So how will I do the novelty detection? Are you saying that I generate something like one single decision boundary across all my training instances, and then based on this decision boundary determine whether a new instance is an outlier or not? | |
| May 21, 2011 at 17:57 | comment | added | Dikran Marsupial | @Andy, sort of, I was suggesting using a one class classifier for novelty detection, but a standard multi-class classifier for performing the classification of observations that are not novelties. I think that may work out better as one-class classifiers are good for novelty detection, but are not likely to be as good as a discriminative classifier for performing the actual classification. Libsvm is an excellent bit of kit, so you are off to a good start. | |
| May 20, 2011 at 19:59 | comment | added | Andy | So if I understand you right, you are basically suggesting doing what I was alluding to in comment (2) in the edits? I just found a one-class SVM implementation in libsvm, so I think I am going to use that as first cut. | |
| May 20, 2011 at 15:25 | history | answered | Dikran Marsupial | CC BY-SA 3.0 |