Currently I am looking for some method to do novelty and outlier detection. I found some good example here using scikit-learn (Link1). However, it is based on supervised learning and I believe the idea behind is the 'One Class SVM'. Therefore, the method requires the training data is not polluted by outliers.

However, I have got some data that contaminated by outliers already. I can't separate a clean dataset for training purpose. I am wondering if there is an unsupervised learning based method to detect the outliers?

For example, my data can be fitted using Mixture of Gaussian model, but before applying GMM, I would like to get rid of the outliers.

Thanks very much. A.

  • 1
    $\begingroup$ Have a look at the ELKI toolkit, which has tons of unsupervised outlier detection algorithms. $\endgroup$ – Has QUIT--Anony-Mousse Dec 4 '14 at 22:04
  • $\begingroup$ @Anony-Mousse Thanks for your reply. Could you name one or two famous unsupervised outlier detection algorithms please? I didn't use Java based ELKI before. Trying to develop or use something in Matlab. $\endgroup$ – Samo Jerom Dec 5 '14 at 10:27
  • 1
    $\begingroup$ DB-Outlier, kNN-Outlier and Local Outlier Factor. Matlab won't scale well for these, because it doesn't support data indexing as far as I know. $\endgroup$ – Has QUIT--Anony-Mousse Dec 5 '14 at 12:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.