Github user dbtsai commented on a diff in the pull request:
https://github.com/apache/spark/pull/11610#discussion_r55963818
--- Diff:
mllib/src/main/scala/org/apache/spark/ml/optim/WeightedLeastSquares.scala ---
@@ -108,6 +101,57 @@ private[ml] class WeightedLeastSquares(
"Consider setting fitIntercept=true.")
}
}
+ /*
+ * If more than one of the features in the data are constant (i.e.
data matrix has constant
+ * columns), then A^T.A is no longer positive definite and Cholesky
decomposition fails
+ * (because the normal equation does not have a solution).
+ * In order to find a solution, we need to drop constant columns from
the data matrix. Or,
+ * we can drop corresponding column and row from A^T.A matrix.
+ * Once we drop rows/columns from A^T.A matrix, the Cholesky
decomposition will produce
+ * correct coefficients. But, for the final result, we need to add
zeros to the list of
+ * coefficients corresponding to the constant features.
+ */
+ val aVarRaw = summary.aVar.values
+ // this will keep track of features to keep in the model, and remove
+ // features with zero variance.
+ val nzVarIndex = aVarRaw.zipWithIndex.filter(_._1 != 0).map(_._2)
+ val nz = nzVarIndex.length
+ // if there are features with zero variance, then ATA is not positive
definite, and we need to
+ // keep track of that.
+ val singular = summary.k > nz
--- End diff --
`isSingular` for readability.
---
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]
