[GitHub] spark issue #14276: [WIP][SPARK-16638][ML][Optimizer] fix L2 reg computation...

[srowen <gi...@git.apache.org>]

Github user srowen commented on the issue:

    https://github.com/apache/spark/pull/14276
  
    It's worth double-checking with @holdenk and @dbtsai. I think this is working as intended since `WeightedLeastSquares` does show multiplying each feature by sigma. To undo it you'd need to divide the partial gradient by its square, and divide the squared coefficient value by its square too in the loss term.
    
    I suppose the logic is that features on a larger scale end up with small coefficients and aren't penalized much in the loss function, so multiplying them by their "scale" compensates. I think this only makes sense when fitting an intercept too, but I haven't thought this through.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastructure@apache.org or file a JIRA ticket
with INFRA.
---

---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org