Github user dbtsai commented on the pull request:
https://github.com/apache/spark/pull/3746#issuecomment-67842962
@bryanyang0528 The learning rate issue here is different story. With modern
optimization algorithms like LBFGS and OWLQN, the learning rate is not
required. The algorithms will find the step size in line search step. As
@srowen pointed out, the statistical property of model will be different
without the 1/2 factor compared with other package. At Alpine Data Labs, I
implemented generalized linear model with elastic net (mixing L1 and L2) using
OWLQN in Spark, I can train and get exactly the same coefficients and the same
statistical property for model including std error, p-value, t-value, residual
plot, and QQ plot, etc. For lots of our customers in financial industry, those
stats are very important, and it's required to get the same solution compared
with well-known R's reference implementation with scalability.
Although I only have limited time on contributing to open source project,
I'll try to have most of my work available in Spark 1.3.
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
