[
https://issues.apache.org/jira/browse/MADLIB-1249?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16530609#comment-16530609
]
Jingyi Mei commented on MADLIB-1249:
------------------------------------
For interface and output, we propose the following:
# use _importance_ param to generate both oob and gini importance: when set to
True, variable oob importance for categorical and continuous features and gini
importance (impurity importance) will be output in the group table
_<model_table>_group._ When set to False, don't print out them in
_<model_table>_group_ table. Default is True.
# The _train_output_group_ table will be look like this:
{code:java}
madlib=# select * from train_output_group;
[ RECORD 1 ]-----------+------------------------------------------------------
gid | 1
success | t
cat_n_levels | {3,2}
cat_levels_in_text | {overcast,rain,sunny,False,True}
oob_error | 0.50000000000000000000
oob_cat_var_importance | {0,0}
oob_con_var_importance | {0,0}
impurity_var_importance | {0,47.7450349919421,17.3997290057814,34.855101856011}
{code}
where impurity_var_importance includes gini importance for both continuous and
categorical variable with the same order of _list_of_features_ param.
3. In essence, gini importance will be calculated for each tree and RF only if
_importance_ is set to be true for better performance.
> Add gini importance to RF
> -------------------------
>
> Key: MADLIB-1249
> URL: https://issues.apache.org/jira/browse/MADLIB-1249
> Project: Apache MADlib
> Issue Type: New Feature
> Components: Module: Random Forest
> Reporter: Jingyi Mei
> Assignee: Jingyi Mei
> Priority: Major
> Fix For: v1.15
>
>
> As a follow up of https://issues.apache.org/jira/browse/MADLIB-1205
> As a data scientist
> I want a measure of variable importance in RF
> so that
> I can understand which predictors are most useful for predicting the response
> variable.
> We can add a similar measure in our RF code and distinguish this from our
> permuted importance metric by calling the current metric as
> {{oob_variable_importance}} and this new metric as
> {{impurity_variable_importance}}.
> Interface
> Details of interface TBD, but involves:
> * RF can use the {{importance}} param to generate both oob and gini
> importance
> * Write out to the {{<output_table_name>_group table}}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
