1
$\begingroup$

I'm doing outlier detection (Conditional Outliers) on a multivariate time series. The outliers appear every 2 weeks $\pm$ 4 days.

How can I incorporate this prior in my models, to reduce the number of false positive? Alternatively, how should I filter the results of my models with this prior knowledge?

All my models use sliding windows and calculate an anomaly score of a series of points.

My rough idea: If an outlier is detected, model a Gaussian distribution where $\mu$ is set 14 days later and $\sigma$ is 4 days wide. Then multiply future predictions with the pmf of the distribution. Assuming the model has some kind of anomaly score, we are reducing the score of points which are unlikely outliers.

$\endgroup$
1
  • $\begingroup$ You can do that. It depends on what approach you are taking, but basically you want a feature to encode 'time since last anomaly'. This allows for drift. Welcome to the site! $\endgroup$
    – Emre
    Jul 11, 2018 at 20:26

1 Answer 1

0
$\begingroup$

One option is to add an explicit rule. In code, the rule would be an if-then conditional statement. This is sometimes called "business logic".

$\endgroup$

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.