Github user facaiy commented on the issue:

    https://github.com/apache/spark/pull/17556
  
    Hi, I has checked R GBM's code and found that:
    R's gbm uses mean value $(x + y) / 2$,  not weighted mean $(c_x * x + c_y * 
y) / (c_x + c_y)$ described in [JIRA 
SPARK-16957](https://issues.apache.org/jira/browse/SPARK-16957), for split 
point.
    
    1. code snippet:
    [gbm-developers/gbm](https://github.com/gbm-developers/gbm)
    commit a1defa382a629f8b97bf9f552dcd821ee7ac9dac
    src/node_search.cpp:145:
    ```c++
          else if(cCurrentVarClasses == 0)   // variable is continuous
          {
            // Evaluate the current split
                dCurrentSplitValue = 0.5*(dLastXValue + dX);
          }
    ```
    
    2. test
    To verify it, I create a toy dataset and take a test on R. 
    ```R
    > f = c(0.0, 0.0, 1.0, 1.0, 1.0, 1.0)
    > l = c(0,   0,   1,   1,   1,   1)
    > df = data.frame(l, f)
    > sapply(df, class)
            l         f
    "numeric" "numeric"
    > mod <- gbm(l~f, data=df, n.trees=1, bag.fraction=1, n.minobsinnode=1, 
distribution = "bernoulli")
    > pretty.gbm.tree(mod)
      SplitVar SplitCodePred LeftNode RightNode MissingNode ErrorReduction 
Weight
    0        0  5.000000e-01        1         2           3       1.333333      
6
    1       -1 -3.000000e-03       -1        -1          -1       0.000000      
2
    2       -1  1.500000e-03       -1        -1          -1       0.000000      
4
    3       -1  1.480297e-19       -1        -1          -1       0.000000      
6
         Prediction
    0  1.480297e-19
    1 -3.000000e-03
    2  1.500000e-03
    3  1.480297e-19
    ```
    As expected,
    the root's split point is 5.000000e-01, namely mean value `0.5 = (0 + 1) / 
2`, not weighted mean `0.66667 = (0 * 2 + 1 * 4) / 6`.
    
    3. conclusion
    I prefer to using weighted mean for split point in the PR, rather than mean 
value in R's gbm package. How about you? @sethah @srowen


---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

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

Reply via email to