3
$\begingroup$

I have a list of accounts as data set and I need to group the accounts that refer to the same user using many features.

I'm thinking to use machine learning( but I'm new in this domain), because I know the group of each account for the training data set.

example of training data:

account-id  Feature1    Feature2    class(Group)
1            T1          P4          Gr1
2            T2          P4          Gr1
3            T3          P2          Gr2

The problem is in the testing of data and when a new account arrive for a new group not learned before in the training set.

example of testing data:

account-id   Feature1   Feature2
4             T5         P5
5             T6         P5
6             T3         P2

The groups of the testing data should be as following:

account-id   Feature1   Feature2   class(Group)
4             T5         P5         Gr3
5             T6         P5         Gr3
6             T3         P2         Gr2

The accounts 4 and 5 are in a new group (Gr3) which is not learned before in the training data.

My question is how could I group the new data under a new class that is not defined before in the learning phase ? and which algorithm can I use to solve this issue ?

$\endgroup$
2
$\begingroup$

There are several ways to do this, I'll just list two:

  1. Train a one versus all classifier $\psi_k$ for each class $k$ in your training set. Predict the test set with each one of those. Test instances that get rejected by each classifier belong to an unseen class.

  2. Train a typical multiclass classifier $\psi_1^\prime$ on your training set and a one-class classifier $\psi_2^\prime$ on all the training data. Test instances that get rejected by the one-class classifier $\psi_2^\prime$ belong to an unseen class, instances that get recognized by $\psi_2^\prime$ can be run through the multiclass classifier $\psi_1^\prime$.

I prefer the first approach, because (i) I feel it gives you more insight into unrecognized test instances and (ii) I'm not a big fan of one-class classification. The former approach also works in case you are facing a multilabel problem.

| cite | improve this answer | |
$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.