WIP: Resolve "Adding new datasets"

code/bolsonaro/models/ensemble_selection_forest_regressor.py

+from sklearn.metrics import mean_squared_error
+from sklearn.base import BaseEstimator
+from sklearn.tree import DecisionTreeRegressor
+from abc import abstractmethod, ABCMeta
+import numpy as np
+from tqdm import tqdm
+
+
+class EnsembleSelectionForestRegressor(BaseEstimator, metaclass=ABCMeta):
+    """
+    'Ensemble selection from libraries of models' by Rich Caruana et al
+    """
+
+    def __init__(self, models_parameters, library, score_metric=mean_squared_error):
+        self._models_parameters = models_parameters
+        self._library = library
+        self._extracted_forest_size = self._models_parameters.extracted_forest_size
+        self._score_metric = score_metric
+
+    @property
+    def models_parameters(self):
+        return self._models_parameters
+
+    @property
+    def library(self):
+        return self._library
+
+    def fit(self, X_train, y_train, X_val, y_val):
+        scores_list = list()
+        for estimator in self._library:
+            val_score = self._score_metric(estimator.predict(X_val), y_val)
+            scores_list.append(val_score)
+
+        class_list = list(self._library)
+        m = np.argmax(np.asarray(scores_list))
+        self._ensemble_selected = [class_list[m]]
+        temp_pred = class_list[m].predict(X_val)
+        del class_list[m]
+        for k in range(self._extracted_forest_size - 1):
+            candidate_index = 0
+            best_score = 100000
+            for j in range(len(class_list)):
+                temp_pred = np.vstack((temp_pred, class_list[j].predict(X_val)))
+                temp_mean = np.mean(temp_pred, axis=0)
+                temp_score = self._score_metric(temp_mean, y_val)
+                if (temp_score < best_score):
+                    candidate_index = j
+                    best_score = temp_score
+                temp_pred = np.delete(temp_pred, -1, 0)
+            self._ensemble_selected.append(class_list[candidate_index])
+            temp_pred = np.vstack((temp_pred, class_list[candidate_index].predict(X_val)))
+            del class_list[candidate_index]
+
+    def score(self, X, y):
+        predictions = self._predict_base_estimator(X)
+        return self._score_metric(predictions, y)
+
+    def predict_base_estimator(self, X):
+        predictions = list()
+        for tree in self._ensemble_selected:
+            predictions.append(tree.predict(X))
+        mean_predictions = np.mean(np.array(predictions), axis=0)
+        return mean_predictions
+
+    @staticmethod
+    def generate_library(X_train, y_train, random_state=None):
+        criterion_arr = ["mse"]#, "friedman_mse", "mae"]
+        splitter_arr = ["best"]#, "random"]
+        depth_arr = [i for i in range(5, 20, 1)]
+        min_samples_split_arr = [i for i in range(2, 20, 1)]
+        min_samples_leaf_arr = [i for i in range(2, 20, 1)]
+        max_features_arr = ["sqrt"]#["auto", "sqrt", "log2"]
+
+        library = list()
+        with tqdm(total=len(criterion_arr) * len(splitter_arr) * \
+            len(depth_arr) * len(min_samples_split_arr) * len(min_samples_leaf_arr) * \
+            len(max_features_arr)) as bar:
+            bar.set_description('Generating library')
+            for criterion in criterion_arr:
+                for splitter in splitter_arr:
+                    for depth in depth_arr:
+                        for min_samples_split in min_samples_split_arr:
+                            for min_samples_leaf in min_samples_leaf_arr:
+                                for max_features in max_features_arr:
+                                    t = DecisionTreeRegressor(criterion=criterion, splitter=splitter, max_depth=depth, min_samples_split=min_samples_split,
+                                        min_samples_leaf=min_samples_leaf, max_features=max_features, random_state=random_state)
+                                    t.fit(X_train, y_train)
+                                    library.append(t)
+                                    bar.update(1)
+        return library