Github user orhankislal commented on a diff in the pull request:
https://github.com/apache/incubator-madlib/pull/142#discussion_r123618828
--- Diff: src/modules/recursive_partitioning/DT_impl.hpp ---
@@ -486,8 +485,18 @@ DecisionTree<Container>::expand(const Accumulator
&state,
Index stats_i = static_cast<Index>(state.stats_lookup(i));
assert(stats_i >= 0);
- // 1. Set the prediction for current node from stats of all
rows
- predictions.row(current) = state.node_stats.row(stats_i);
+ if (statCount(predictions.row(current)) !=
+ statCount(state.node_stats.row(stats_i))){
+ // Predictions for each node is set by its parent using
stats
+ // recorded while training parent node. These stats do not
include
+ // rows that had a NULL value for the primary split
feature.
+ // The NULL count is included in the 'node_stats' while
training
+ // current node. Further, presence of NULL rows indicate
that
+ // stats used for deciding 'children_wont_split' are
inaccurate.
+ // Hence avoid using the flag to decide termination.
+ predictions.row(current) = state.node_stats.row(stats_i);
+ children_wont_split = false;
+ }
--- End diff --
Let me try to rephrase my question:
Assume the if check at lines 490-491 is `true`. `children_wont_split` will
be set to `false`. This variable is used with an & operator at line 568. That
means the second boolean value is irrelevant, the result will always be
`false`. In this case, do we need the 2 double for loops from lines 516-547?
The compiler might calculate the backwards slice of this instruction to
find its optimal location but I don't think we should rely on that.
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
