How do I implement one class SVM for user authentication? I have unlabelled dataset for 100 users. In my understanding, I have to create 100 different one class svm models for all users to determine if the user is authentic or not , but I am not sure about it.
Hi Anu,
In general this approach is for outlier detection. However, to apply it, you need to have *only* one target label for all the instances (100 users) that you have. In other words all the 100 users have to share one class label (doing that manually is a common approach).
HTH.
Samer Sarsam, PhD.
Thank you for your response.
I would like to elaborate my question. My problem is to distinguish an authentic user from unauthentic user. So I am using one class svm so that it could distinguish one user from the other. I have dataset for 100 users.
My experimental setup is as follows:
For user1:
Train data: 80% data of User1
Test data: 20% of data of User1 + some data of User2 or any other user other than user1
For user2:
Train data: 80% data of User2
Test data: 20% of data of User2 + some data of User3 or any other user other than user2
and so on for all 100 users.
I am evaluation model in terms of EER for each user.
Am I in the correct track? Any suggestions would be appreciated.
No worries.
I recommend you not to use one-class classification for your task, but instead of that use binary classification. To do so, create one dataset (training set) for all the users and label each user as an authentic/unauthentic user. After that, clean your data well and use either k-fold cross-valuation or holdout method to evaluate your classification algorithm. Finally, examine the performance of your classification algorithm using accuracy, ROC, etc.
Cheers,
Samer Sarsam, PhD.
Thank you so much again.
But I am still not clear on the part you said: "create one dataset (training set) for all the users and label each user as an authentic/unauthentic user ".
I am trying to learn behaviour of users from dataset and based on that determine if user is authentic or not. In my understanding, all users as authentic.
For example for User1,
- all data that is present in User1 dataset is normal for user1
- all data of User2 or any other user is outlier.
So how do you think I can label all users as authentic or not authentic? This might sound stupid question but I am new in this area and any help will be of great value! :)
- Badges
- Science topic