[ 
https://issues.apache.org/jira/browse/SPARK-30641?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17307690#comment-17307690
 ] 

Weichen Xu commented on SPARK-30641:
------------------------------------

Good work!

> Project Matrix: Linear Models revisit and refactor
> --------------------------------------------------
>
>                 Key: SPARK-30641
>                 URL: https://issues.apache.org/jira/browse/SPARK-30641
>             Project: Spark
>          Issue Type: New Feature
>          Components: ML, PySpark
>    Affects Versions: 3.1.0, 3.2.0
>            Reporter: zhengruifeng
>            Priority: Major
>
> We had been refactoring linear models for a long time, and there still are 
> some works in the future. After some discuss among [~huaxingao] [~srowen] 
> [~weichenxu123] , we decide to gather related works under a sub-project 
> Matrix, it include:
>  # *Blockification (vectorization of vectors)*
>  ** vectors are stacked into matrices, so that high-level BLAS can be used 
> for better performance. (about ~3x faster on sparse datasets, up to ~18x 
> faster on dense datasets, see SPARK-31783 for details).
>  ** Since 3.1.1, LoR/SVC/LiR/AFT supports blockification, and we need to 
> blockify KMeans in the future.
>  # *Standardization (virutal centering)*
>  ** Existing impl of standardization in linear models does NOT center the 
> vectors by removing the means, for the purpose of keeping dataset 
> _*sparsity*_. However, this will cause feature values with small var be 
> scaled to large values, and underlying solver like LBFGS can not efficiently 
> handle this case. see SPARK-34448 for details.
>  ** If internal vectors are centers (like other famous impl, i.e. 
> GLMNET/Scikit-Learn), the convergence ratio will be better. In the case in 
> SPARK-34448, the number of iteration to convergence will be reduced from 93 
> to 6. Moreover, the final solution is much more close to the one in GLMNET.
>  ** Luckily, we find a new way to _*virtually*_ center the vectors without 
> densifying the dataset. Good results had been observed in LoR, we will take 
> it into account in other linear models.
>  # _*Initialization (To be discussed)*_
>  ** Initializing model coef with a given model, should be beneficial to: 1, 
> convergence ratio (should reduce number of iterations); 2, model stability 
> (may obtain a new solution more close to the previous one);
>  # _*Early Stopping* *(To be discussed)*_
>  ** we can compute the test error in the procedure (like tree models), and 
> stop the training procedure if test error begin to increase;
>  
>       If you want to add other features in these models, please comment in 
> the ticket.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org

Reply via email to