From 49aa494813b57f691274a2d9327cb9cb3fe00649 Mon Sep 17 00:00:00 2001
From: Julie Tibshirani <jtibs@cs.stanford.edu>
Date: Tue, 20 Aug 2019 22:54:06 -0700
Subject: [PATCH] Ensure quantile forests give consistent estimates across
 platforms.

We now include a tie-breaker while sorting the list of sample IDs and outcomes.
Without the tie-breaker, the sort could produce a different element ordering
across different runs of the algorithm.
---
 .../prediction/QuantilePredictionStrategy.cpp  |  6 +++++-
 .../resources/quantile_oob_predictions.csv     | 18 +++++++++---------
 2 files changed, 14 insertions(+), 10 deletions(-)

diff --git a/core/src/prediction/QuantilePredictionStrategy.cpp b/core/src/prediction/QuantilePredictionStrategy.cpp
index 511126495..d9556bb90 100644
--- a/core/src/prediction/QuantilePredictionStrategy.cpp
+++ b/core/src/prediction/QuantilePredictionStrategy.cpp
@@ -52,7 +52,11 @@ std::vector<double> QuantilePredictionStrategy::compute_quantile_cutoffs(
   std::sort(samples_and_values.begin(),
             samples_and_values.end(),
             [](std::pair<size_t, double> first_pair, std::pair<size_t, double> second_pair) {
-              return first_pair.second < second_pair.second;
+              // Note: we add a tie-breaker here to ensure that this sort consistently produces the
+              // same element ordering. Otherwise, different runs of the algorithm could result in
+              // different quantile predictions on the same data.
+              return first_pair.second < second_pair.second
+                  || (first_pair.second == second_pair.second && first_pair.first < second_pair.first);
             });
 
   std::vector<double> quantile_cutoffs;
diff --git a/core/test/forest/resources/quantile_oob_predictions.csv b/core/test/forest/resources/quantile_oob_predictions.csv
index da10cfe6b..534c7d97f 100644
--- a/core/test/forest/resources/quantile_oob_predictions.csv
+++ b/core/test/forest/resources/quantile_oob_predictions.csv
@@ -121,7 +121,7 @@
 -10, -10, 10
 -10, -10, 10
 -1, -1, 1
--1, -1, 1
+-1, 1, 1
 -1, -1, 1
 10, 10, 10
 -10, -10, -10
@@ -410,7 +410,7 @@
 -1, -1, 1
 -1, -1, 1
 -1, -1, 1
--10, -10, 10
+-10, 10, 10
 -1, -1, 1
 10, 10, 10
 -1, -1, 1
@@ -421,7 +421,7 @@
 -10, 10, 10
 -1, -1, 1
 -1, -1, 1
--1, -1, 1
+-1, 1, 1
 -1, -1, 1
 -1, 1, 1
 -10, 10, 10
@@ -459,7 +459,7 @@
 -1, -1, 1
 -10, 10, 10
 -10, -10, 10
--10, 10, 10
+-10, -10, 10
 -1, -1, 1
 -1, -1, 1
 -10, -1, 10
@@ -537,7 +537,7 @@
 -10, 10, 10
 -1, 1, 1
 -1, -1, 1
--10, 10, 10
+-10, -10, 10
 -10, -10, 10
 -1, 1, 1
 -1, 1, 1
@@ -589,7 +589,7 @@
 -1, 1, 1
 -1, -1, 1
 -10, 10, 10
--1, -1, 1
+-1, 1, 1
 -1, 1, 1
 -1, -1, 1
 -10, -10, 10
@@ -832,7 +832,7 @@
 -10, -10, 10
 -1, 1, 1
 -10, -10, -10
--1, 1, 1
+-1, -1, 1
 -1, 1, 1
 -1, -1, 1
 -10, -10, 10
@@ -847,7 +847,7 @@
 -1, -1, 1
 -1, -1, 1
 -1, -1, 1
--10, -10, 10
+-10, 10, 10
 -10, -10, 10
 -10, 10, 10
 -1, -1, 1
@@ -857,7 +857,7 @@
 10, 10, 10
 -10, -10, 1
 -10, 10, 10
--10, -1, 10
+-10, -10, 10
 -1, -1, -1
 -1, 1, 1
 -1, 1, 1