Hi Ady, Overfitting is a possible explanation. If your model learnt your normal scenarios too well then every abnormal data will be predicted as abnormal (so you will have a good performance for anomalies) however none of the normal instances of the test set will be in the normal region (so you will have a high FPR).
Albert On Wed, 5 Apr 2017 at 15:37, Ady Wahyudi Paundu <awpau...@gmail.com> wrote: > Good day Scikit-Learn Masters, > > I have used Scikit-Learns OCSVM module previously with satisfying results. > However on my current tasks I have this problem for one-class analysis: > > In my previous cases, I used OCSVM for Anomaly detector, and the > normal classes in each cases were coming from one scenario. > Now, I want to create one Anomaly detector system, with multiple > normal scenario (in this case, 3 different normal scenario). Lets say > I have scenario A, B and C, and I want to distinguish all data that is > not coming from A and B and C. > What I have been tried is combining all training data A and B and C > into one data set and fit it using OCSVM module. When I tested the > output model to several anomaly data-set it worked good. However, when > I tested it against either one of the normal scenario, it gave a very > high False Positives (AUROC: 99%). > > So my question, is it because a bad approach? by combining all the > different normal data set into one training data set. > Or is it because I was using it (the OCSVM) wrong? (i use 'rbf' kernel > with nu and gamma set to 0.001) > Or is it the case with wrong tools? another algorithm perhaps? > > I dont know if this is a proper question to ask here, so if it is not > (maybe because this is just a Machine Learning question in general), > just disregard it. > > Thank you in advance > > Best regards, > Ady > _______________________________________________ > scikit-learn mailing list > scikit-learn@python.org > https://mail.python.org/mailman/listinfo/scikit-learn >
_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn
