24788: Adds utilization of multithreading when computing stats for nominals, misc fixes

howso/react_aggregate.amlg

@@ -118,11 +118,11 @@
 			confusion_matrix_min_count 15
 			;{type "list" values "string"}
 			;features for which to determine classes or values that cause the highest accuracy and/or prediction contribution to the
-			;specified 'value_robust_contributions_action_features'
+			;specified 'value_robust_contributions_action_feature'
 			value_robust_contributions_features (null)
 			;{type "string"}
 			;action feature being predicted whose accuracy and prediction affects should be computed by 'value_robust_contributions_features'
-			value_robust_contributions_action_features (null)
+			value_robust_contributions_action_feature (null)
 			;{type "number" exclusive_min 0}
 			;number of maximum buckets to bin continuous values into, defaults to 30 when unspecified
 			value_robust_contributions_num_buckets 30
@@ -200,10 +200,10 @@
 			;		context_num_samples: optional, limit on the number of context cases when 'context_condition_precision' is set to 'similar'.
 			;			If null, will be set to k. default is null
 			;		value_robust_accuracy_contributions: optional, none/true/false.
-			;			Perform a focused computation to determine how all the individual values of specified 'value_robust_contributions_features' affect the accuracy of 'value_robust_contributions_action_features'.
+			;			Perform a focused computation to determine how all the individual values of specified 'value_robust_contributions_features' affect the accuracy of 'value_robust_contributions_action_feature'.
 			;			Outputs under 'value_robust_contributions'
 			;		value_robust_prediction_contributions: optional, none/true/false.
-			;			Perform a focused computation to determine how all the individual values of specified 'value_robust_contributions_features' affect the predictions  of 'value_robust_contributions_action_features'.
+			;			Perform a focused computation to determine how all the individual values of specified 'value_robust_contributions_features' affect the predictions  of 'value_robust_contributions_action_feature'.
 			;			Outputs under 'value_robust_contributions'
 			;	)
 			details (null)
@@ -638,7 +638,7 @@
 					))
 				)
 			value_robust_contributions_case_ids
-				(if (or (get details "value_robust_accuracy_contributions") (get details "value_robust_prediction_contributions)"))
+				(if (or (get details "value_robust_accuracy_contributions") (get details "value_robust_prediction_contributions"))
 					(call !SampleCases (assoc
 						num
 							(if num_robust_accuracy_contributions_samples
@@ -1108,7 +1108,7 @@
 								weight_feature weight_feature
 								hyperparam_map custom_hyperparam_map
 								context_condition_filter_query context_condition_filter_query
-								action_feature value_robust_contributions_action_features
+								action_feature value_robust_contributions_action_feature
 								value_robust_contributions_features value_robust_contributions_features
 								max_num_buckets value_robust_contributions_num_buckets
 								valid_weight_feature valid_weight_feature

howso/residuals.amlg

@@ -188,27 +188,45 @@
 	;					i.e.: { "residual_map" : { "featureA" : 3.4... } }
 	; using_shared_deviations: boolean, default (null). If true, will only use parent features and expand according to the shared deviations groups
 	#!ExpandResidualValuesToUncertainty
-	(set
-		feature_residuals_map
-		"residual_map"
-		(let (assoc
-			temp_residual_map
-				(map
+	(declare
+		(assoc
+			has_shared_deviations (and using_shared_deviations (size !sharedDeviationsMap))
+		)
+
+		;check if any feature residuals need to be modified
+		(declare (assoc
+			need_modification_residuals_map
+				(filter (map
 					(lambda
 						;inactive features are bound to min residual
 						(if (contains_index !inactiveFeaturesMap (current_index))
 							(get !cachedFeatureMinResidualMap (current_index))
 
-							;set upper bound to nominals to be max nominal deviation, accounting for imbalanced classes
+							;if residual is too small, replace it with the minimim allowed
+							(<
+								(current_value)
+								(max
+									(get !cachedFeatureMinResidualMap (current_index))
+									(get !userSpecifiedFeatureErrorsMap (current_index))
+								)
+							)
+							(max
+								(get !cachedFeatureMinResidualMap (current_index))
+								(get !userSpecifiedFeatureErrorsMap (current_index))
+							)
+
+							(= (null) (current_value))
+							(max (get !cachedFeatureMinResidualMap (current_index)) (get !userSpecifiedFeatureErrorsMap (current_index)) )
+
+							;else nominal may need to set upper bound, accounting for imbalanced classes
 							(contains_index !nominalsMap (current_index))
 							(let
 								(assoc
 									;map of class -> count
 									feature_grouping
-										(if (size !sharedDeviationsMap)
+										(if has_shared_deviations
 											(call !GetSharedDeviationGrouping (assoc feature_group_to_retrieve (current_index 2)))
 										)
-									current_feature (current_index 1)
 								)
 								(declare (assoc
 									class_counts_map
@@ -229,89 +247,99 @@
 												;list of each feature's class count assocs
 												(map
 													(lambda
-														(compute_on_contained_entities
-															(query_value_masses (current_value) (null))
-														)
+														(compute_on_contained_entities (query_value_masses (current_value) (null)) )
 													)
 													feature_grouping
 												)
 											)
-
-											(compute_on_contained_entities
-												(query_value_masses current_feature (null))
-											)
+											;else a map of class -> count
+											(compute_on_contained_entities (query_value_masses (current_index 1) (null)) )
 										)
-								))
 
-								;do not max cap nominal deviations if less than 2 classes have been trained
-								(if (<= (size class_counts_map) 1)
-									(max
-										(get !cachedFeatureMinResidualMap (current_index))
-										(current_value)
-										(get !userSpecifiedFeatureErrorsMap (current_index))
-									)
+								))
 
-									;else cap the max nominal deviation
+								;only need to cap max nominal deviations if 2 or more classes have been trained
+								(if (> (size class_counts_map) 1)
 									(let
-										(assoc total_count (apply "+" (values class_counts_map)) )
-
-										(min
+										(assoc
+											total_count (apply "+" (values class_counts_map))
+										)
+										(declare (assoc
 											;nominal max deviation is the sum of: each class's probability multiplied by probability of getting it wrong
-											(apply "+"
-												(map
-													(lambda (let
-														(assoc class_prob (* (/ (current_value 1) total_count)) )
-														(* class_prob (- 1 class_prob))
-													))
-													(values class_counts_map)
+											nominal_max_deviation
+												(apply "+"
+													(map
+														(lambda (let
+															(assoc class_prob (/ (current_value 1) total_count) )
+															(* class_prob (- 1 class_prob))
+														))
+														(values class_counts_map)
+													)
 												)
-											)
-
-											(max
-												(get !cachedFeatureMinResidualMap (current_index))
-												(current_value)
-												(get !userSpecifiedFeatureErrorsMap (current_index))
-											)
+										))
+										(if (> (current_value) nominal_max_deviation)
+											nominal_max_deviation
 										)
 									)
 								)
 							)
-
-							;else continuous value don't have upper bounds, set the lower bound
-							(max
-								(get !cachedFeatureMinResidualMap (current_index))
-								(current_value)
-								(get !userSpecifiedFeatureErrorsMap (current_index))
-							)
 						)
 					)
-
 					;if using shared deviations, map over a reduced residuals map with only features that are not
 					; in a shared deviations group or are the primary keys of a shared deviations group
-					(if (and
-							using_shared_deviations
-							(size !sharedDeviationsMap)
-						)
+					(if has_shared_deviations
 						(remove
 							(get feature_residuals_map "residual_map")
 							!sharedDeviationsNonPrimaryFeatures
 						)
 
 						(get feature_residuals_map "residual_map")
 					)
-				)
+				))
+		))
+
+		(if has_shared_deviations
+			(set
+				feature_residuals_map
+				"residual_map"
+				(call !ExpandForSharedDeviations (assoc
+					compressed_values
+						(if (size need_modification_residuals_map)
+							;iterate over the residual map and use the values from need_modification_residuals_map where those values are defined
+							(map
+								(lambda
+									(if (!= (null) (last (current_value))) (last (current_value)) (first (current_value)) )
+								)
+								(remove (get feature_residuals_map "residual_map") !sharedDeviationsNonPrimaryFeatures)
+								need_modification_residuals_map
+							)
+
+							(remove (get feature_residuals_map "residual_map") !sharedDeviationsNonPrimaryFeatures)
+						)
+				))
 			)
 
-			(if (and
-					using_shared_deviations
-					(size !sharedDeviationsMap)
+			;else output the modified residuals
+			(size need_modification_residuals_map)
+			(set
+				feature_residuals_map
+				"residual_map"
+				;iterate over the residual map and use the values from need_modification_residuals_map where those values are defined
+				(map
+					(lambda
+						(if (!= (null) (last (current_value))) (last (current_value)) (first (current_value)) )
+					)
+					(get feature_residuals_map "residual_map")
+					need_modification_residuals_map
 				)
-				(call !ExpandForSharedDeviations (assoc compressed_values temp_residual_map))
-				temp_residual_map
 			)
+
+			;else output unmodified residuals
+			feature_residuals_map
 		)
 	)
 
+
 	;helper method to sample cases from the dataset, outputs a list of case ids
 	;if there are more cases than the sample size, randomly select that many cases, by default cases are in random order
 	#!SelectCaseIdsForResiduals

howso/train.amlg

@@ -1218,8 +1218,8 @@
 							session (get_value session)
 							session_training_index
 								(if (size ts_ablated_indices_map)
-									(+ next_trained_index (get ts_ablated_indices_map (+ input_case_index (current_value 1))))
-									(+ next_trained_index (current_value 1))
+									(+ session_training_index (get ts_ablated_indices_map (+ input_case_index (current_value 1))))
+									(+ session_training_index (current_value 1))
 								)
 						))
 					))

howso/train_utilities.amlg

@@ -289,14 +289,16 @@
 			(lambda (let
 				(assoc
 					feature_values
-						(if encode_features_on_train
-							(call !ConvertFromInput (assoc
-								feature_values (unzip (current_value 2) rebalance_feature_indices)
-								features continuous_rebalance_features
-							))
+						(if (size continuous_rebalance_features)
+							(if encode_features_on_train
+								(call !ConvertFromInput (assoc
+									feature_values (unzip (current_value 2) rebalance_feature_indices)
+									features continuous_rebalance_features
+								))
 
-							;else use feature_values as-is
-							(unzip (current_value 1) rebalance_feature_indices)
+								;else use feature_values as-is
+								(unzip (current_value 1) rebalance_feature_indices)
+							)
 						)
 					;default nominal weight to 1
 					nominal_rebalance_features_weight 1
@@ -337,6 +339,7 @@
 					(let
 						(assoc
 							nominal_values
+								(if (size nominal_rebalance_feature_indices)
 									(if encode_features_on_train
 										(call !ConvertFromInput (assoc
 											feature_values (unzip (current_value 2) nominal_rebalance_feature_indices)
@@ -346,6 +349,7 @@
 										;else use feature_values as-is
 										(unzip (current_value 1) nominal_rebalance_feature_indices)
 									)
+								)
 						)
 
 						(assign (assoc

howso/types.amlg

@@ -1510,7 +1510,7 @@
 								description
 									(concat
 										"Perform a focused computation to determine how all the individual values of specified "
-										"'value_robust_contributions_features' affect the accuracy of 'value_robust_contributions_action_features'."
+										"'value_robust_contributions_features' affect the accuracy of 'value_robust_contributions_action_feature'."
 									)
 							)
 						value_robust_prediction_contributions
@@ -1519,7 +1519,7 @@
 								description
 									(concat
 										"Perform a focused computation to determine how all the individual values of specified "
-										"'value_robust_contributions_features' affect the predictions of 'value_robust_contributions_action_features'."
+										"'value_robust_contributions_features' affect the predictions of 'value_robust_contributions_action_feature'."
 									)
 							)
 					)

unit_tests/ut_h_rebalance_features.amlg

@@ -124,9 +124,9 @@
 				[ 3 0.75 	1 ]
 				;class 'M' reciprocal is 1/(4+1)=0.2, new total rebalance weight is 2.5333,
 				;new total mass is 8, thus scalar is 8 / 2.5333 = 3.158,  0.2 * 3.158 = 0.6316
-				;distributing about 0.6316 worth of weight among two cases with a ~ 30/70 split:
-				[ 4 0.93 	1.3 ]
-				[ 5 1.19 	1.7 ]
+				;distributing about 0.6316 worth of weight among two cases with a ~ 25/75 split:
+				[ 4 0.91 	1.25 ]
+				[ 5 1.22 	1.75 ]
 				[ 6 1		1 ]
 			]
 	))
@@ -156,8 +156,8 @@
 				[ 1 1.5 	1 ]
 				[ 2 0.75 	1 ]
 				[ 3 0.75 	1 ]
-				[ 4 0.93 	1.3 ]
-				[ 5 1.19 	1.7 ]
+				[ 4 0.91 	1.25 ]
+				[ 5 1.22 	1.75 ]
 				[ 6 1 	 	1 ]
 				;unknown class 'L' value is (0.5+0.25)/2=0.375, new total rebalance weight is 2.908
 				;new total mass is 9, thus scalar is 9 / 2.908 = 3.095, 0.375 * 3.095 = 1.16