Description
A/B Online testing is widely used to compare online models. We can implement A/B testing in MLlib and integrate it with Spark Streaming. For example, we have a PairDStream[String, Double], whose keys are model ids and values are observations (click or not, or revenue associated with the event). With A/B testing, we can tell whether one model is significantly better than another at a certain time. There are some caveats. For example, we should avoid multiple testing and support A/A testing as a sanity check.
Attachments
Issue Links
- is related to
-
SPARK-11960 User guide section for streaming a/b testing
- Resolved
- links to