GitHub user iyounus opened a pull request:
https://github.com/apache/spark/pull/11610
[SPARK-13777] [ML] Remove constant features from training in noraml solver
(WLS)
## What changes were proposed in this pull request?
"normal" solver in LinearRegression uses Cholesky decomposition to
calculate the coefficients. If the data has features with identical values
(zero variance), then (A^T A) matrix is not positive definite any more and the
Cholesky decomposition fails.
Since A^T.A and features variances are calculated in single pass, it's
better to modify ATA instead to re-calculating it from the data after dropping
constant columns. In this PR, I'm dropping columns and rows from ATA
corresponding to features with zero variance. Then the cholesky decomposition
can be performed without any problem.
## How was this patch tested?
A unit test under LineatReagessionSuite is added which compares results
from this change and l-bgfs solver to glmnet. All these are now onsistent.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/iyounus/spark
SPARK-13777_WLS_fix_for_constant_features
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/spark/pull/11610.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #11610
----
commit 9412ef439d523e58e5d6b8294628c2196f6f9019
Author: Imran Younus <[email protected]>
Date: 2016-03-09T19:38:12Z
remove constant features from training in noraml solver
----
---
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]
