[ https://issues.apache.org/jira/browse/SPARK-13777?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15187791#comment-15187791 ]
Imran Younus commented on SPARK-13777: -------------------------------------- My understanding is that the Cholesky decomposition method requires decomposition of A^T.A matrix instead of A. One can calculate A^T.A in single pass through the data and move it the driver, where the decomposition can be done. As far as I know, this cannot be done with QR decomposition method of solving normal eqaution. > Weighted Leaset Squares fails when there are features with identical values. > ---------------------------------------------------------------------------- > > Key: SPARK-13777 > URL: https://issues.apache.org/jira/browse/SPARK-13777 > Project: Spark > Issue Type: Bug > Components: ML > Reporter: Imran Younus > Priority: Minor > > "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. > For the same case, "l-bfgs" solver sets the coefficients of these constant > features to zero and produces valid coefficients for the rest of the > features. This behaviour is consistent with glmnet in R. "normal" solver > should also do the same. -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org