4
$\begingroup$

Concept drift is when the relation between the input data and the target variable changes over time. like changes in the conditional distribution.

is novelty an outlier? what should I think of? what is the difference between concept drift and novelty and anomaly? is the concept drift considered a type of novelty? how exactly? can you please explain !!

$\endgroup$
2
$\begingroup$

Roughly all three concepts are related.

Drift means the relationship between input and output is dynamic and changes (stochastically) over (sufficiently long periods of) time. That is, it is not stationary. For example, consumers' criteria about what to buy, change over time, for example as people become more eco-conscious. More importantly drift, when it happens, invalidates the existing model used for prediction.

Anomaly also called an outlier is a very rare non-typical event (when input-output relationship is considered stationary over time), that happens upon exceptional circumstances. Something like a white snake. It may happen but is not typical of snakes and if it happens it does not mean that input-output relationship has necessarily drifted from the original assumptions (eg the assumptions about the color distribution of snakes). Accordingly anomaly, when it happens, does not invalidate the existing model used for prediction.

Novelty as far as I understand it, is an umbrella term for something new and unpredictable happening, which however may be attributable to anything (drift, anomaly, etc).

Please note that determining the reason for the observed novelty requires careful analysis (for example, multiple anomalies may mean drift is what is actually happening)!

References:

  1. Anomaly detection
  2. Concept drift
$\endgroup$
4
  • $\begingroup$ thanks for your fully explained answer but what about novelty does it also invalidates the existing model used for prediction? and can I think of the concept as a type of novelty? under what conditions? $\endgroup$ Oct 14 at 19:43
  • $\begingroup$ Novelty term, per this answer can mean any unpredictable event, including drift. So yes drift is a type of novelty. Whether novelty invalidates the model depends on the type of novelty, an outlier novelty, does not invalidate the model. $\endgroup$
    – Nikos M.
    Oct 14 at 19:51
  • $\begingroup$ in this case how to make the model adapt to such change whether it's novelty type or concept drift? $\endgroup$ Oct 14 at 20:21
  • $\begingroup$ This is the holy grail of machine learning. There is not clear cut answer, only some approaches that might work in some cases. Read the references. $\endgroup$
    – Nikos M.
    Oct 14 at 21:40

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.