Current implementation of SplitHonestExamples is not following the descriptions in "Generalized Random Forests" (Athey et al.).
|
void SplitHonestExamples( |
|
const absl::Span<const UnsignedExampleIdx> selected_examples, |
|
const float leaf_rate, utils::RandomEngine* random_engine, |
|
std::vector<UnsignedExampleIdx>& leaf_examples, |
|
std::vector<UnsignedExampleIdx>& working_selected_examples) { |
|
std::uniform_real_distribution<float> dist_01; |
|
|
|
// Reduce the risk of std::vector re-allocations. |
|
const float error_margin = 1.1f; |
|
leaf_examples.reserve(selected_examples.size() * leaf_rate * error_margin); |
|
working_selected_examples.reserve(selected_examples.size() * |
|
(1.f - leaf_rate) * error_margin); |
|
|
|
for (const auto& example : selected_examples) { |
|
if (dist_01(*random_engine) < leaf_rate) { |
|
leaf_examples.push_back(example); |
|
} else { |
|
working_selected_examples.push_back(example); |
|
} |
|
} |
|
} |
In the captions of the GRF paper Algorithm 1, it specifies that:
SplitSample randomly divides a set into two evenly-sized, non-overlapping halves.
However, the current YDF honesty implementation is not accounting for this property.
