1
$\begingroup$

concept drift usually refers to the change in the relationship between input and output data over time.

I do have dataset of users' activity in an e-commerce website. Let's say we have a sequence of item-view actions (a user session). I can label each action by some concept.

E.g. we have a sequence of item views such as:

Item ID: 1, 2, 3, 4, 5

let's label it with some concept (category, tag, topic or whatever), so we get this:

Concept: tv, tv, tv, headphones, headphones

So a user changed their interest from tv to headphones.

Is this still a concept drift? We can say that no change in underlying mapping function has happened. Therefore it does not meet the usual definition of what the concept drift is.

$\endgroup$

1 Answer 1

0
$\begingroup$

As you correctly say, this is not a case of concept drift.
Concept drift means that the statistical properties of the concept one is trying to predict have change. This can mean changes in $P(X,Y)$ in a supervised context but also just in $P(X)$ if your task is unsupevised. Further, concept drift is a phenomenon observed in a sample, but not necessarily on the level of individual observations. Here your example is easily misleading, because your data is of sequential nature - you observe a given number of users over time and get multiple measurements per user over this time frame.
Considering such a sample, we can observe that for example $P(\mathbf{x}|y_1)$ or $P(y_1)$ has changed and this could indicate the presence of concept drift (but might aswell just be connected to the sampling process).
Just one user in your example looking at headphones instead of tv's does not necessarily consitute such a change, as it has to be viewed on the level of the entire sample.

$\endgroup$
2
  • $\begingroup$ Can you suggest me what should I look for, if I want to study changes of concept/interest within a sequence, as showed in question? I have observed (by manual browsing) patterns in changes of concept in sequence which are typical for a concept drift (sudden, gradual etc). I want to do some more research on how to quantify it (for analysis) and how to help my NN for sequence prediction make better predictions by taking such patterns and their characteristics into an account. $\endgroup$
    – miro
    Nov 13, 2019 at 9:35
  • $\begingroup$ I figure that would be worth a separate question $\endgroup$
    – deemel
    Nov 15, 2019 at 8:14

Your Answer

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

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