1
$\begingroup$

I have developed online neural network based one-class classifier and also enabled it for forgetting mechanism. So, now online learning with forgetting mechanism is possible. But how to handle if data is non-stationary. As it is one-class classification so, suppose we have two class data i.e. normal and outlier class then trained by only normal class data in online fashion. But training data might have different distribution compared to distribution of normal and outlier data in testing as their distribution are changing. So, how algorithm will handle this drift/shift.

One more doubt regarding non-stationary data handling: whether we will pass trained labelled data once and further labelled data for training are collected from unlabelled testing data?

$\endgroup$
2
  • $\begingroup$ I think your question is an interesting one. But first you have to give more details bef anyone can help. First, what is this one-class classification? is the response a binary variable where 1 is the identified class, 0 otherwise. Pls provide details how your forgetting mechanism works. When you say it's "trained by normal data", do you mean your training features are (jointly) normally distributed? $\endgroup$
    – horaceT
    May 6, 2017 at 4:01
  • $\begingroup$ Thanks for the reaponse. I hope I made question clear now. Normal means simply a class type/name normal. It is no linked with normal distribution. Forgetting mechaanism is just an unlearning part which is required in any non-stationary data handling algo as it simply unlearn the model from old samples and make them learn from new samples. But my doubt is how will u decide which samples to select for training from testing samples or from somewhere else. $\endgroup$ May 6, 2017 at 9:48

1 Answer 1

2
$\begingroup$

Usually, non-stationary series can be dealt by differencing the time series once or twice. You can add verification to your tool - checking auto-correlation plot (or partial correlation plot).

$\endgroup$
2
  • $\begingroup$ Whether we will pass trained labelled data once and further labelled data for training are collected from unlabelled testing data or we need to pass the labelled data manually or through sliding window. $\endgroup$ May 7, 2017 at 13:45
  • $\begingroup$ Good question, try both actually $\endgroup$ May 9, 2017 at 18:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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