add option flags for shuffling feature order

Author: paulbkoch

docs/benchmarks/ebm-benchmark.ipynb

@@ -34,8 +34,6 @@
    "source": [
     "# install powerlift if not already installed\n",
     "\n",
-    "# !! IMPORTANT !! : until the next release, install locally with \"pip install -e .[datasets,postgres]\" from powerlift directory\n",
-    "\n",
     "try:\n",
     "    import powerlift\n",
     "except ModuleNotFoundError:\n",

python/interpret-core/interpret/develop.py

@@ -6,8 +6,15 @@
 _current_module = sys.modules[__name__]
 _current_module.is_debug_mode = False
 
+# Global options
 _purify_boosting = False
 _purify_result = False
+_randomize_initial_feature_order = True
+# TODO: investigate if _randomize_feature_order actually decreases accuracy
+# https://github.com/interpretml/interpret/issues/563#issuecomment-2240820952
+# this seems to decrease accuracy slightly, but helps with collinearity
+_randomize_greedy_feature_order = True  # randomize feature order only if greedy enabled
+_randomize_feature_order = False  # randomize feature order always
 
 
 def print_debug_info(file=None):

python/interpret-core/interpret/glassbox/_ebm/_boost.py

@@ -7,6 +7,7 @@
 
 import heapq
 from ...utils._native import Native
+from ... import develop
 
 import logging
 
@@ -74,7 +75,7 @@ def boost(
             _log.info("Start boosting")
             native = Native.get_native_singleton()
             nominals = native.extract_nominals(dataset)
-            random_cyclic_ordering = np.empty(len(term_features), np.int64)
+            random_cyclic_ordering = np.arange(len(term_features), dtype=np.int64)
 
             while step_idx < max_steps:
                 term_boost_flags_local = term_boost_flags
@@ -85,7 +86,13 @@ def boost(
                         bestkey = None
                         heap = []
                         # if pure cyclical then only randomize at start
-                        if 0 < greedy_steps or step_idx == 0:
+                        if (
+                            step_idx == 0
+                            and develop._randomize_initial_feature_order
+                            or develop._randomize_greedy_feature_order
+                            and 0 < greedy_steps
+                            or develop._randomize_feature_order
+                        ):
                             # TODO: test if shuffling during pure cyclic is better
                             native.shuffle(rng, random_cyclic_ordering)