[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14377205#comment-14377205 ] Xiangrui Meng commented on SPARK-3278: -- Did you try truncating the digits of x to reduce the number of possible buckets? If the loss of precision is not super important, this could help scalability. Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Fix For: 1.3.0 Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14362986#comment-14362986 ] Martin Zapletal commented on SPARK-3278: Vladimir, just to update you on the progress. I was able to complete the isotonic regression with 100M records, but failed with insufficient memory error with 150M records on my machine. You may be able to run with larger amounts of data on better machines. Pool adjacent violators algorithm can theoretically have linear time complexity, but although I have used the best algorithm I could find I am not convinced it reaches this efficiency. I will work on providing evidence. The biggest issue with the current algorithm is however with the parallelization approach. Its properties are unfortunately nowhere near linear scalability (linear solution time increase with linear parallelism increase or constant solution time with linear parallelism increase and linear problem size increase). This was expected and is caused by the algorithm itself for the following reasons 1) The algorithm works in two steps. First the computation is distributed to all partitions, but then gathered and the algorithm is run again on the whole data set. This approach may leave most of work for the last sequential step and thus gaining very little compared to purely sequential implementation or even performing worse. That can happen in case where parallel isotonic regressions return a locally optimal solution that will however have to change for a global solution in the last step. Another performance drawback in comparison to sequential processing is the potential need to copy data to each process. 2) It requires the whole dataset to fit into one process’ memory in the last step (or repeated disk access). I started looking into the issue and was able to design an iterative algorithm that adressed both the above issues and performed very close to linear scalability. It however still has correctness (rounding) issues and will require further research. Let me know if that helped. In the meantime I will continue working on benchmarks and performance quantification of the current algorithm as well as on research for potentially more efficient solutions. Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Fix For: 1.3.0 Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14355119#comment-14355119 ] Vladimir Vladimirov commented on SPARK-3278: Martin. This would be really nice. Is it possible to run isotonic regression on the data I'll provide? (~150-200M records). The answers I'm looking for - how long it would take to train the model on this data set, how much resources it would take on a cluster and confirm that it won't blow spark. I'll export values in format float1, float2 per line - similar to how it is described in the doc http://people.apache.org/~pwendell/spark-1.3.0-rc1-docs/mllib-isotonic-regression.html ? Where float1 - is between 0 and 1. And float2 - is 0 or 1 Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Fix For: 1.3.0 Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14353252#comment-14353252 ] Vladimir Vladimirov commented on SPARK-3278: Had anyone benchmarked the performance of Spark Isotonic Regression implementation on big datasets (100 M, 1000M) ? Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Fix For: 1.3.0 Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14353549#comment-14353549 ] Martin Zapletal commented on SPARK-3278: What particular benchmarks would you like to see? I can do them. Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Fix For: 1.3.0 Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14353419#comment-14353419 ] Xiangrui Meng commented on SPARK-3278: -- I don't know any. It really depends on how may buckets it outputs. I can imagine problems with 100M buckets. Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Fix For: 1.3.0 Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14229271#comment-14229271 ] Apache Spark commented on SPARK-3278: - User 'zapletal-martin' has created a pull request for this issue: https://github.com/apache/spark/pull/3519 Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Assignee: Martin Zapletal Add isotonic regression for score calibration. -- 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
[jira] [Commented] (SPARK-3278) Isotonic regression
[ https://issues.apache.org/jira/browse/SPARK-3278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14182064#comment-14182064 ] Martin Zapletal commented on SPARK-3278: I am interested in working on this ticket. Can you please assign to me? Isotonic regression --- Key: SPARK-3278 URL: https://issues.apache.org/jira/browse/SPARK-3278 Project: Spark Issue Type: New Feature Components: MLlib Reporter: Xiangrui Meng Add isotonic regression for score calibration. -- 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