update quantile computation in quantilegbm and quantilelgbm

configs/ref_gradientboost_quantiles.yaml

@@ -56,7 +56,7 @@ learning:
       'min_weight_fraction_leaf': Real(0.0, 0.5, prior='uniform')
 
 prediction:
-  quantiles: 0.95
+  quantiles: 0.9
   outbands: 4
 
 

configs/ref_quantile_lgbm.yaml

@@ -38,11 +38,13 @@ learning:
     target_transform: identity
     random_state: 1
     max_depth: 20
+    upper_alpha: 0.95
+    lower_alpha: 0.05
 
 
 prediction:
   prediction_template: configs/data/sirsam/dem_foc2.tif
-  quantiles: 0.95
+  quantiles: 0.90
   outbands: 4
 
 

uncoverml/optimise/models.py

@@ -641,6 +641,7 @@ def __init__(self, target_transform='identity',
         self.alpha = alpha
         self.upper_alpha = upper_alpha
         self.lower_alpha = lower_alpha
+        self.interval = upper_alpha - lower_alpha
 
     @staticmethod
     def collect_prediction(regressor, X_test):
@@ -658,7 +659,7 @@ def fit(self, X, y, *args, **kwargs):
     def predict(self, X, *args, **kwargs):
         return self.predict_dist(X, *args, **kwargs)[0]
 
-    def predict_dist(self, X, interval=0.95, *args, ** kwargs):
+    def predict_dist(self, X, interval=0.90, *args, ** kwargs):
         Ey = self.gb.predict(X)
 
         ql_ = self.collect_prediction(self.gb_quantile_lower, X)
@@ -667,9 +668,17 @@ def predict_dist(self, X, interval=0.95, *args, ** kwargs):
         Vy = ((qu_ - ql_) / (norm.ppf(self.upper_alpha) - norm.ppf(self.lower_alpha))) ** 2
 
         # to make gbm quantile model consistent with other quantile based models
-        ql, qu = norm.interval(interval, loc=Ey, scale=np.sqrt(Vy))
-
-        return Ey, Vy, ql, qu
+        if interval == self.interval:
+            return Ey, Vy, ql_, qu_
+        else:
+            # if the interval matches (upper_alpha-lower_alpha), we don't need to compute ql, qu
+            # and also don't need to make assumpition of normal distribution to compute ql and qu
+            log.warn("===============================================")
+            log.warn("Used normal distribution assumption to compute quantiles."
+                     " Using quantiles=(upper_alpha-lower_alpha) will remove this requirement!")
+            log.warn("===============================================")
+            ql, qu = norm.interval(self.interval, loc=Ey, scale=np.sqrt(Vy))
+            return Ey, Vy, ql, qu
 
 
 class GBMReg(GradientBoostingRegressor, TagsMixin):
@@ -820,7 +829,7 @@ def fit(self, X, y, *args, **kwargs):
     def predict(self, X, *args, **kwargs):
         return self.predict_dist(X, *args, **kwargs)[0]
 
-    def predict_dist(self, X, interval=0.95, *args, ** kwargs):
+    def predict_dist(self, X, interval=0.9, *args, ** kwargs):
         Ey = self.gb.predict(X)
 
         ql_ = self.collect_prediction(self.gb_quantile_lower, X)
@@ -829,9 +838,17 @@ def predict_dist(self, X, interval=0.95, *args, ** kwargs):
         Vy = ((qu_ - ql_) / (norm.ppf(self.upper_alpha) - norm.ppf(self.lower_alpha))) ** 2
 
         # to make gbm quantile model consistent with other quantile based models
-        ql, qu = norm.interval(interval, loc=Ey, scale=np.sqrt(Vy))
-
-        return Ey, Vy, ql, qu
+        if interval == self.interval:
+            return Ey, Vy, ql_, qu_
+        else:
+            # if the interval matches (upper_alpha-lower_alpha), we don't need to compute ql, qu
+            # and also don't need to make assumpition of normal distribution to compute ql and qu
+            log.warn("===============================================")
+            log.warn("Used normal distribution assumption to compute quantiles."
+                     " Using quantiles=(upper_alpha-lower_alpha) will remove this requirement!")
+            log.warn("===============================================")
+            ql, qu = norm.interval(self.interval, loc=Ey, scale=np.sqrt(Vy))
+            return Ey, V, ql, qu
 
 
 class CatBoostWrapper(CatBoostRegressor, TagsMixin):