Realtime Anomaly Detection

User 5284 | 6/12/2016, 10:16:08 PM

Hello -

Does GraphLab have a functionality to do realtime anomaly detection that can handle seasonality ? The specific feature I am looking for is similar to https://github.com/twitter/AnomalyDetection. Would love to try if you have such a feature and can provide me pointers. Thanks!

Karthik

Comments

User 12 | 6/14/2016, 5:18:18 PM

Hi Karthik, GraphLab Create's anomaly detection toolkit does have a moving z-score model that can handle seasonality if the moving window size is set carefully.

https://dato.com/learn/userguide/anomalydetection/movingzscore.html

The challenge with Twitter's ESD model (based on a season trend decomposition with loess smoothers), is that users have to know the seasonal period with high confidence beforehand. In my experience, the moving Z-score is a bit more forgiving when this number is not known. That said, we're always looking for feedback and requests for new tools, so please feel free to DM me if you can share more details about your use case.

Thanks! Brian


User 5284 | 6/25/2016, 12:10:28 AM

Thanks Brian. Send DM.