How to test unsupervised learning methods for anomaly detection? I am looking for a test strategy to evaluate my result of my anomaly detection technique? what is your offer more than evaluate with different algorithms. My data is some time series very low frequency.
$\begingroup$
$\endgroup$
1
asked Jun 25, 2021 at 14:37
-
1$\begingroup$ You might read this paper "How to Evaluate the Quality of Unsupervised Anomaly Detection Algorithms?" => arxiv.org/abs/1607.01152 $\endgroup$– user119783Jun 25, 2021 at 18:32
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Each observation will have a "position" in your problem's feature-space.
- If your data is already clustered, calculate the distance to the nearest cluster centers, and see how it compares to the average distance of transactions associated with those clusters.
- If your data is not clustered, grab a random sample of data, calculate the distance to the N nearest neighbors; as new data comes in, look for cases where distance to nearest neighbors is significantly larger than average.
Ideally, the dimensions of feature-space should be normalized, so that each dimension of interest spans a similar length.
-
$\begingroup$ Thanks but there are huge amount of data, and as a result, it does not make sense to use kmeans $\endgroup$ Jul 14, 2021 at 15:59