Adding SimpleImputer, IterativeImputer and KNNImputer to the config space.

tpot2/config/get_configspace.py

@@ -45,7 +45,8 @@
 from sklearn.feature_selection import f_classif, f_regression #TODO create a selectomixin using these?
 from sklearn.discriminant_analysis import LinearDiscriminantAnalysis, QuadraticDiscriminantAnalysis
 from sklearn.gaussian_process import GaussianProcessRegressor, GaussianProcessClassifier
-from sklearn.impute import SimpleImputer
+from sklearn.experimental import enable_iterative_imputer
+from sklearn.impute import SimpleImputer, IterativeImputer, KNNImputer
 
 all_methods = [SGDClassifier, RandomForestClassifier, ExtraTreesClassifier, GradientBoostingClassifier, MLPClassifier, DecisionTreeClassifier, XGBClassifier, KNeighborsClassifier, SVC, LogisticRegression, LGBMClassifier, LinearSVC, GaussianNB, BernoulliNB, MultinomialNB, ExtraTreesRegressor, RandomForestRegressor, GradientBoostingRegressor, BaggingRegressor, DecisionTreeRegressor, KNeighborsRegressor, XGBRegressor,  ZeroCount, ColumnOneHotEncoder, Binarizer, FastICA, FeatureAgglomeration, MaxAbsScaler, MinMaxScaler, Normalizer, Nystroem, PCA, PolynomialFeatures, RBFSampler, RobustScaler, StandardScaler, SelectFwe, SelectPercentile, VarianceThreshold, SGDRegressor, Ridge, Lasso, ElasticNet, Lars, LassoLars, LassoLarsCV, RidgeCV, SVR, LinearSVR, AdaBoostRegressor, GradientBoostingRegressor, RandomForestRegressor, BaggingRegressor, ExtraTreesRegressor, DecisionTreeRegressor, KNeighborsRegressor, ElasticNetCV,
                AdaBoostClassifier,MLPRegressor,
@@ -56,7 +57,7 @@
                GaussianProcessClassifier, BaggingClassifier,LGBMRegressor,
                Passthrough,SkipTransformer,
                PassKBinsDiscretizer,
-               SimpleImputer,
+               SimpleImputer, IterativeImputer, KNNImputer
                ]
 
 
@@ -124,7 +125,7 @@
         "all_transformers" : ["transformers", "scalers"],
 
         "arithmatic": ["AddTransformer", "mul_neg_1_Transformer", "MulTransformer", "SafeReciprocalTransformer", "EQTransformer", "NETransformer", "GETransformer", "GTTransformer", "LETransformer", "LTTransformer", "MinTransformer", "MaxTransformer"],
-        "imputers": ["SimpleImputer"],
+        "imputers": ["SimpleImputer", "IterativeImputer", "KNNImputer"],
         "skrebate": ["ReliefF", "SURF", "SURFstar", "MultiSURF"],
         "genetic_encoders": ["DominantEncoder", "RecessiveEncoder", "HeterosisEncoder", "UnderDominanceEncoder", "OverDominanceEncoder"],
 
@@ -136,8 +137,6 @@
 
 def get_configspace(name, n_classes=3, n_samples=1000, n_features=100, random_state=None):
     match name:
-        case "SimpleImputer":
-            return imputers.simple_imputer_cs
 
         #autoqtl_builtins.py
         case "FeatureEncodingFrequencySelector":
@@ -352,6 +351,12 @@ def get_configspace(name, n_classes=3, n_samples=1000, n_features=100, random_st
             ) 
 
         #imputers.py
+        case "SimpleImputer":
+            return imputers.simple_imputer_cs
+        case "IterativeImputer":
+            return imputers.get_IterativeImputer_config_space(n_features=n_features, random_state=random_state)
+        case "KNNImputer":
+            return imputers.get_KNNImputer_config_space(n_samples=n_samples)
 
         #mdr_configs.py
         case "MDR":
@@ -443,8 +448,17 @@ def get_node(name, n_classes=3, n_samples=100, n_features=100, random_state=None
         sfm_sp = get_configspace(name="SelectFromModel", n_classes=n_classes, n_samples=n_samples, random_state=random_state)
         ext = get_node("ExtraTreesRegressor", n_classes=n_classes, n_samples=n_samples, random_state=random_state)
         return WrapperPipeline(estimator_search_space=ext, method=SelectFromModel, space=sfm_sp)
-
+    # TODO Add IterativeImputer with more estimator methods
+    '''
+    if name == "IterativeImputer_learnedestimators":
+        iteative_sp = get_configspace(name="IterativeImputer", n_classes=n_classes, n_samples=n_samples, random_state=random_state)
+        regessor_searchspace = get_search_space(["LinearRegression", ..], n_classes=n_classes, n_samples=n_samples, random_state=random_state)
+        return WrapperPipeline(estimator_search_space=regressor_searchspace, method=ItartiveImputer, space=iteative_sp)
+    '''
     #these are nodes that have special search spaces which require custom parsing of the hyperparameters
+    if name == "IterativeImputer":
+        configspace = get_configspace(name, n_classes=n_classes, n_samples=n_samples, random_state=random_state)
+        return EstimatorNode(STRING_TO_CLASS[name], configspace, hyperparameter_parser=imputers.IterativeImputer_hyperparameter_parser)
     if name == "RobustScaler":
         configspace = get_configspace(name, n_classes=n_classes, n_samples=n_samples, random_state=random_state)
         return EstimatorNode(STRING_TO_CLASS[name], configspace, hyperparameter_parser=transformers.robust_scaler_hyperparameter_parser)

tpot2/config/imputers.py

@@ -1,9 +1,80 @@
+import sklearn
+import sklearn.ensemble
+import sklearn.linear_model
+import sklearn.neighbors
 from ConfigSpace import ConfigurationSpace
 from ConfigSpace import ConfigurationSpace, Integer, Float, Categorical, Normal
+from ConfigSpace import EqualsCondition
+
 
 simple_imputer_cs = ConfigurationSpace(
     space = {
-        'strategy' : Categorical('strategy', ['mean','median', 'most_frequent', ]),
-        'add_indicator' : Categorical('add_indicator', [True, False]), 
+        'strategy' : Categorical('strategy', 
+                                 ['mean','median', 'most_frequent', 'constant']
+                                 ),
+        #'add_indicator' : Categorical('add_indicator', [True, False]), 
+        #Removed add_indicator, it appends a mask next to the rest of the data 
+        # and can cause errors. gk
+    }
+)
+
+def get_IterativeImputer_config_space(n_features, random_state):
+    space = { 'initial_strategy' : Categorical('initial_strategy', 
+                                             ['mean', 'median', 
+                                              'most_frequent', 'constant']),
+                'n_nearest_features' : Integer('n_nearest_features', 
+                                           bounds=(1, n_features)),
+                'imputation_order' : Categorical('imputation_order', 
+                                             ['ascending', 'descending', 
+                                              'roman', 'arabic', 'random']),
     }
-)
+
+    estimator = Categorical('estimator', ['Bayesian', 'RFR', 'Ridge', 'KNN'])  
+    sample_posterior = Categorical('sample_posterior', [True, False])
+    sampling_condition = EqualsCondition(sample_posterior, estimator, 'Bayesian')
+
+    if random_state is not None: 
+            #This is required because configspace doesn't allow None as a value
+            space['random_state'] = random_state
+
+    cs = ConfigurationSpace(space=space)
+    cs.add_hyperparameters([estimator, sample_posterior])
+    cs.add_conditions([sampling_condition])
+    return cs
+
+def get_KNNImputer_config_space(n_samples):
+    space = {
+            'n_neighbors': Integer('n_neighbors', bounds=(1, max(n_samples,100))),
+            'weights': Categorical('weights', ['uniform', 'distance'])
+    }
+
+    return ConfigurationSpace(
+          space=space
+    )
+
+def IterativeImputer_hyperparameter_parser(params):
+    est = params['estimator']
+    match est:
+        case 'Bayesian':
+                estimator = sklearn.linear_model.BayesianRidge()
+        case 'RFR':
+                estimator = sklearn.ensemble.RandomForestRegressor()
+        case 'Ridge':
+                estimator = sklearn.linear_model.Ridge()
+        case 'KNN':
+                estimator = sklearn.neighbors.KNeighborsRegressor()
+
+    final_params = {
+            'estimator' : estimator,
+            'initial_strategy' : params['initial_strategy'],
+            'n_nearest_features' : params['n_nearest_features'],
+            'imputation_order' : params['imputation_order'],
+    }
+
+    if 'sample_posterior' in params:
+        final_params['sample_posterior'] = params['sample_posterior']
+
+    if 'random_state' in params:
+        final_params['random_state'] = params['random_state']
+
+    return final_params