Streaming detection drifts when people are left out of the loop.
Thresholds drift, definitions vary by operator, and feedback matters. Outlier detection in real environments is rarely a one shot batch job.
-
“Can it keep up with a stream?”
Moving windows
Deep Isolation Forest trains on a recent window and scores the next one, to emulate data arriving over time.
-
“What if the model is wrong?”
Human feedback in the loop
Analysts relabel, add or delete points, and those edits feed into later runs.
-
“Who decides what counts as an outlier?”
Thresholds you can tune
Fixed thresholds, or a relative rule from the current score distribution, adjusted in the interface.
Monitoring gets stronger when human corrections are part of the work.
What it proves.
Operator loop design for model monitoring.
- Additional experiments compared Deep Isolation Forest against classical detectors (EIF, iForest, LOF, KNN) using static evaluation workflows.
- For categorical streams, the backend can keep one model per category and merge the results into one stream.
An honest note
The Deep Isolation Forest algorithm itself is imported from an external module. This project integrates it rather than implementing it from scratch.
Why it matters for your AI
This operator loop maps cleanly to agent trace review and drift workflows.
See the code
Tell me where your AI needs a person in the loop.
The first 30 minute call is free, and I reply within one business day.