Decision Tree with Polynomial Basis Functions

scripts/forecast_model.py

@@ -15,33 +15,44 @@
 warnings.filterwarnings("ignore", category=FitFailedWarning)
 
 
-def forecast_dec2019():
+def forecast_dec2019(two_features=False):
     ''' forecast model for Dec 2019 eruption
     '''
     # constants
     month = timedelta(days=365.25/12)
     day = timedelta(days=1)
     td = TremorData()
-
-    # construct model object
-    data_streams = ['rsam','mf','hf','dsar']
-    fm = ForecastModel(ti='2011-01-01', tf='2020-01-01', window=2., overlap=0.75, 
-        look_forward=2., data_streams=data_streams)
-
-    # columns to manually drop from feature matrix because they are highly correlated to other 
-    # linear regressors
-    drop_features = ['linear_trend_timewise','agg_linear_trend']
-
+
     # set the available CPUs higher or lower as appropriate
     n_jobs = 6
 
+    # construct model object
+    if two_features:
+        data_streams = ['rsam']
+        fm = ForecastModel(ti='2012-04-01', tf='2012-10-01', window=2., overlap=0.75,
+                           look_forward=2., data_streams=data_streams, root='twoFeatures')
+        use_only_features = ['rsam__maximum', 'rsam__fft_coefficient__coeff_12__attr_"abs"']
+        classifier = 'DTPBF'
+        drop_features = []
+    else:
+        data_streams = ['rsam','mf','hf','dsar']
+        fm = ForecastModel(ti='2011-01-01', tf='2020-01-01', window=2., overlap=0.75,
+            look_forward=2., data_streams=data_streams)
+
+        # columns to manually drop from feature matrix because they are highly correlated to other
+        # linear regressors
+        drop_features = ['linear_trend_timewise', 'agg_linear_trend']
+        use_only_features = []
+        classifier = "DT"
+
     # train the model, excluding 2019 eruption
-    # note: building the feature matrix may take several hours, but only has to be done once 
+    # note: building the feature matrix may take several hours, but only has to be done once
     # and will intermittantly save progress in ../features/
     # trained scikit-learn models will be saved to ../models/*root*/
     te = td.tes[-1]
-    fm.train(ti='2011-01-01', tf='2020-01-01', drop_features=drop_features, retrain=True, 
-        exclude_dates=[[te-month,te+month],], n_jobs=n_jobs)      
+    fm.train(ti='2011-01-01', tf='2020-01-01', drop_features=drop_features, retrain=True,
+             exclude_dates=[[te-month,te+month],], n_jobs=n_jobs,
+             classifier=classifier, use_only_features=use_only_features)
 
     # run forecast from 2011 to 2020
     # model predictions will be saved to ../predictions/*root*/ 
@@ -58,24 +69,31 @@ def forecast_dec2019():
     fm.hires_forecast(ti=te-fm.dtw-fm.dtf, tf=te+month/30, recalculate=True, 
         save=r'{:s}/forecast_hires.png'.format(fm.plotdir), n_jobs=n_jobs)
 
-def forecast_test():
+def forecast_test(two_features=False):
     ''' test scale forecast model
     '''
     # constants
     month = timedelta(days=365.25/12)
-
-    # set up model
-    data_streams = ['rsam','mf','hf','dsar']
-    fm = ForecastModel(ti='2012-04-01', tf='2012-10-01', window=2., overlap=0.75, 
-        look_forward=2., data_streams=data_streams, root='test')
-
+
     # set the available CPUs higher or lower as appropriate
     n_jobs = 6
-
-    # train the model
-    drop_features = ['linear_trend_timewise','agg_linear_trend']
-    fm.train(ti='2012-04-01', tf='2012-10-01', drop_features=drop_features, retrain=True,
-        n_jobs=n_jobs)      
+
+    # set up model
+    if two_features:
+        data_streams = ['rsam']
+        fm = ForecastModel(ti='2012-04-01', tf='2012-10-01', window=2., overlap=0.75,
+                           look_forward=2., data_streams=data_streams, root='testPF')
+        fm.train(ti='2012-04-01', tf='2012-10-01', retrain=True,
+                 n_jobs=n_jobs, classifier='DTPBF',
+                 use_only_features=['rsam__maximum', 'rsam__fft_coefficient__coeff_12__attr_"abs"'])
+    else:
+        data_streams = ['rsam', 'mf', 'hf', 'dsar']
+        fm = ForecastModel(ti='2012-04-01', tf='2012-10-01', window=2., overlap=0.75,
+                           look_forward=2., data_streams=data_streams, root='test')
+        # train the model
+        drop_features = ['linear_trend_timewise', 'agg_linear_trend']
+        fm.train(ti='2012-04-01', tf='2012-10-01', drop_features=drop_features, retrain=True,
+                 n_jobs=n_jobs)
 
     # plot a forecast for a future eruption
     te = fm.data.tes[1]
@@ -115,7 +133,7 @@ def forecast_now():
         save='current_forecast.png', nztimezone=True, n_jobs=n_jobs)  
 
 if __name__ == "__main__":
-    #forecast_dec2019()
-    forecast_test()
+    #forecast_dec2019(two_features=True)
+    forecast_test(two_features=True)
     #forecast_now()
 

whakaari/__init__.py

@@ -55,9 +55,11 @@
 from sklearn.naive_bayes import GaussianNB
 from sklearn.linear_model import LogisticRegression
 from sklearn.svm import SVC
+from sklearn.pipeline import Pipeline, make_pipeline
+from sklearn.preprocessing import PolynomialFeatures
 
 datas = ['rsam','mf','hf','dsar']
-all_classifiers = ["SVM","KNN",'DT','RF','NN','NB','LR']
+all_classifiers = ["SVM","KNN",'DT', 'DTPBF', 'RF','NN','NB','LR']
 _MONTH = timedelta(days=365.25/12)
 _DAY = timedelta(days=1.)
 
@@ -969,6 +971,7 @@ def train(self, ti=None, tf=None, Nfts=20, Ncl=100, retrain=False, classifier="D
             SVM - Support Vector Machine.
             KNN - k-Nearest Neighbors
             DT - Decision Tree
+            DTPBF - Decision Tree with Polynomial Basis Functions
             RF - Random Forest
             NN - Neural Network
             NB - Naive Bayes
@@ -1590,6 +1593,7 @@ def get_classifier(classifier):
         SVM - Support Vector Machine.
         KNN - k-Nearest Neighbors
         DT - Decision Tree
+        DTPBF - Decision Tree with Polynomial Basis Functions
         RF - Random Forest
         NN - Neural Network
         NB - Naive Bayes
@@ -1607,6 +1611,13 @@ def get_classifier(classifier):
         model = DecisionTreeClassifier(class_weight='balanced')
         grid = {'max_depth': [3,5,7], 'criterion': ['gini','entropy'],
             'max_features': ['auto','sqrt','log2',None]}
+    elif classifier == 'DTPBF':
+        model = Pipeline([('polynomial', PolynomialFeatures()),
+                          ('clf', DecisionTreeClassifier())
+                          ])
+        grid = {'clf__max_depth': [3, 5, 7], 'clf__criterion': ['gini', 'entropy'],
+                'clf__max_features': ['auto', 'sqrt', 'log2', None],
+                'polynomial__degree': [1, 2, 3, 4, 5]}
     elif classifier == "RF":        # random forest
         model = RandomForestClassifier(class_weight='balanced')
         grid = {'n_estimators': [10,30,100], 'max_depth': [3,5,7], 'criterion': ['gini','entropy'],