Merge pull request #36 from loft-br/feat/refactor-extrapolation

Feat/refactor extrapolation

Author: GabrielGimenez

README.md

@@ -217,7 +217,7 @@ from xgbse.extrapolation import extrapolate_constant_risk
 survival = bootstrap_estimator.predict(X_valid)
 
 # extrapolating
-survival_ext = extrapolate_constant_risk(survival, 450, 11)
+survival_ext = extrapolate_constant_risk(survival, 450, 15)
 ```
 
 <img src="img/extrapolation.png">
@@ -407,7 +407,7 @@ To cite this repository:
   author = {Davi Vieira and Gabriel Gimenez and Guilherme Marmerola and Vitor Estima},
   title = {XGBoost Survival Embeddings: improving statistical properties of XGBoost survival analysis implementation},
   url = {http://github.com/loft-br/xgboost-survival-embeddings},
-  version = {0.2.1},
+  version = {0.2.2},
   year = {2021},
 }
 ```

docs/examples/extrapolation_example.md

@@ -383,7 +383,7 @@ Notice that this predicted survival curve does not end at zero (cure fraction du
 from xgbse.extrapolation import extrapolate_constant_risk
 
 # extrapolating predicted survival
-survival_ext = extrapolate_constant_risk(survival, 450, 11)
+survival_ext = extrapolate_constant_risk(survival, 450, 15)
 survival_ext.head()
 ```
 

examples/extrapolation_example.ipynb

@@ -34,7 +34,7 @@
 
 setuptools.setup(
     name="xgbse",
-    version="0.2.1",
+    version="0.2.2",
     author="Loft Data Science Team",
     author_email="bandits@loft.com.br",
     description="Improving XGBoost survival analysis with embeddings and debiased estimators",

img/extrapolation.png

@@ -42,9 +42,11 @@
 )
 
 preds = xgbse_model.predict(X_test)
+interval = 10
+final_time = max(time_bins) + 1000
 n_windows = 100
-final_time = max(T_train) + 1000
-preds_ext = extrapolate_constant_risk(preds, final_time=final_time, n_windows=n_windows)
+
+preds_ext = extrapolate_constant_risk(preds, final_time=final_time, intervals=interval)
 
 
 def extrapolation_shape():

setup.py

@@ -6,7 +6,7 @@
 from ._meta import XGBSEBootstrapEstimator
 
 
-__version__ = "0.2.1"
+__version__ = "0.2.2"
 
 __all__ = [
     "XGBSEDebiasedBCE",

tests/test_extrapolation.py

@@ -9,7 +9,7 @@
 
 # lib utils
 from xgbse._base import XGBSEBaseEstimator, DummyLogisticRegression
-from xgbse.converters import convert_data_to_xgb_format, convert_y
+from xgbse.converters import convert_data_to_xgb_format, convert_y, hazard_to_survival
 
 # at which percentiles will the KM predict
 from xgbse.non_parametric import get_time_bins, calculate_interval_failures
@@ -346,9 +346,7 @@ def _predict_from_lr_list(self, lr_estimators, leaves_encoded, time_bins):
 
         # converting these interval predictions
         # to cumulative survival curve
-        preds = (1 - preds).cumprod(axis=1)
-
-        return preds
+        return hazard_to_survival(preds)
 
     def predict(self, X, return_interval_probs=False):
         """

xgbse/__init__.py

@@ -112,3 +112,16 @@ def build_xgb_cox_dmatrix(X, T, E):
     target = np.where(E, T, -T)
 
     return xgb.DMatrix(X, label=target)
+
+
+def hazard_to_survival(interval):
+    """Convert hazards (interval probabilities of event) into survival curve
+
+    Args:
+        interval ([pd.DataFrame, np.array]): hazards (interval probabilities of event)
+        usually result of predict or  result from _get_point_probs_from_survival
+
+    Returns:
+        [pd.DataFrame, np.array]: survival curve
+    """
+    return (1 - interval).cumprod(axis=1)

xgbse/_debiased_bce.py

@@ -1,9 +1,10 @@
 import numpy as np
 import pandas as pd
 from xgbse.non_parametric import _get_conditional_probs_from_survival
+from xgbse.converters import hazard_to_survival
 
 
-def extrapolate_constant_risk(survival, final_time, n_windows, lags=-1):
+def extrapolate_constant_risk(survival, final_time, intervals, lags=-1):
     """
     Extrapolate a survival curve assuming constant risk.
 
@@ -13,7 +14,7 @@ def extrapolate_constant_risk(survival, final_time, n_windows, lags=-1):
 
         final_time (Float): Final time for extrapolation
 
-        n_windows (Int): Number of time windows to compute from last time window in survival to final_time
+        intervals (Int): Time in each interval between last time in survival dataframe and final time
 
         lags (Int): Lags to compute constant risk.
             if negative, will use the last "lags" values
@@ -24,28 +25,23 @@ def extrapolate_constant_risk(survival, final_time, n_windows, lags=-1):
         pd.DataFrame: Survival dataset with appended extrapolated windows
     """
 
-    # calculating conditionals and risk at each time window
-    conditionals = _get_conditional_probs_from_survival(survival)
-    window_risk = 1 - conditionals
+    last_time = survival.columns[-1]
+    # creating windows for extrapolation
+    # here we sum intervals in times to exclude the last time, that already is in surv dataframe and
+    #  to include final time in resulting dataframe
+    extrap_windows = np.arange(last_time + intervals, final_time + intervals, intervals)
 
-    # calculating window sizes
-    time_bins = window_risk.columns.to_series()
-    window_sizes = time_bins - time_bins.shift(1).fillna(0)
+    # calculating conditionals and hazard at each time window
+    hazards = _get_conditional_probs_from_survival(survival)
 
-    # using window sizes to calculate risk per unit time and average risk
-    risk_per_unit_time = np.power(window_risk, 1 / window_sizes)
-    average_risk = risk_per_unit_time.iloc[:, lags:].mean(axis=1)
+    # calculating avg hazard for desired lags
+    constant_haz = hazards.values[:, lags:].mean(axis=1).reshape(-1, 1)
 
-    # creating windows for extrapolation
-    last_time = survival.columns[-1]
-    extrap_windows = np.linspace(last_time, final_time, n_windows) - last_time
+    # repeat hazard for n_windows required
+    constant_haz = np.tile(constant_haz, len(extrap_windows))
 
-    # loop for extrapolated windows
-    for delta_t in extrap_windows:
+    constant_haz = pd.DataFrame(constant_haz, columns=extrap_windows)
 
-        # running constant risk extrapolation
-        extrap_survival = np.power(average_risk, delta_t) * survival.iloc[:, -1]
-        extrap_survival = pd.Series(extrap_survival, name=last_time + delta_t)
-        survival = pd.concat([survival, extrap_survival], axis=1)
+    hazards = pd.concat([hazards, constant_haz], axis=1)
 
-    return survival
+    return hazard_to_survival(hazards)