I want to know if I can use the k-means clustering algorithm for a one class classification (as in the case of one class SVM), which means I have data for 2 classes, and I labelled only the one class that I used for training?
$\begingroup$
$\endgroup$
4
-
$\begingroup$ Can you clarify what means that you only labelled one? $\endgroup$– Juan Esteban de la CalleApr 22 '19 at 2:07
-
$\begingroup$ knowing that I have 2 classes(positive/negative), I mean that I'm doing the training phase with only positive classe (labeled data) and while testing I'll use the two classes. I want the model could detect the negative data. $\endgroup$– KahinaApr 22 '19 at 2:49
-
$\begingroup$ Why don't you have the two classes in both sets? $\endgroup$– Juan Esteban de la CalleApr 22 '19 at 2:50
-
$\begingroup$ the project , I'm working on, aims to test a behavioral approach, which means learning only positive behavior to detect negative one. I should test more than one classifier. I have tested it using one-class SVM, but I'm not sure that it's possible using kMC because it's unsupervised learning algorithm. $\endgroup$– KahinaApr 22 '19 at 2:54
Add a comment
|
$\begingroup$
$\endgroup$
The K-means algorithm has the capacity of retrieving which are the "boundaries" your data has for knowing the only-class, is possible that you don't find the only-class boundaries to be the same boundaries your k-means algorithm found. This is the risk of comparing k-means with the one class classification: Clustering can be looking different things from one class classification.
This answer may guide you to another solution.
How to perform model selection for One-Class Classification?
answered Apr 22 '19 at 2:15
$\begingroup$
$\endgroup$
A Gaussian Mixture Model can be seen as a generalization of k-means, with soft (probabilistic) cluster assignments rather than hard ones. It can be, and often is, used in a one-class setting. It allows for multiple Gaussian components and for non-spherical shapes of clusters, which can be an advantage with many datasets. But if one restricts to single components with spherical covariance matrix, it is very similar to k-means.
