patching some metrics and errors

xgbse/metrics.py

@@ -147,7 +147,7 @@ def approx_brier_score(y_true, survival, aggregate="mean"):
     # adding censoring distribution survival at event
     event_time_windows = _match_times_to_windows(times, survival.columns)
     scoring_df["cens_at_event"] = censoring_dist[event_time_windows].iloc[0].values
-    # TODO Something is broken when use sklearn.model_selection.cross_val_score
+
     # list of window results
     window_results = []
 
@@ -177,11 +177,15 @@ def approx_brier_score(y_true, survival, aggregate="mean"):
         )
 
         # adding and taking average
-        result = (first_term + second_term).sum() / scoring_df.shape[0]
+        # OLD CODE:
+        # result = (first_term + second_term).sum() / scoring_df.shape[0]
+
+        added_terms = (first_term + second_term)
+        result = np.nanmean(added_terms[np.isfinite(added_terms)])
         window_results.append(result)
 
     if aggregate == "mean":
-        return np.array(window_results).mean()
+        return np.nanmean(window_results)
     elif aggregate is None:
         return np.array(window_results)
     else:

xgbse/non_parametric.py

@@ -10,10 +10,11 @@ def get_time_bins(T, E, size=12):
     """
     Method to automatically define time bins
     """
-
-    lower_bound = max(T[E == 0].min(), T[E == 1].min()) + 1
-    upper_bound = min(T[E == 0].max(), T[E == 1].max()) - 1
-
+    # OLD CODE:
+    # lower_bound = max(T[E == 0].min(), T[E == 1].min()) + 1
+    # upper_bound = min(T[E == 0].max(), T[E == 1].max()) - 1
+    lower_bound = np.nanmin(T[E == 1]) + 1
+    upper_bound = np.nanmax(T[E == 1]) - 1
     return np.linspace(lower_bound, upper_bound, size, dtype=int)
 
 
@@ -74,12 +75,19 @@ def sample_time_bins(surv_array, T_neighs, time_bins):
     """
 
     surv_df = []
-
     for t in time_bins:
-        survival_at_t = (surv_array + (T_neighs > t)).min(axis=1)
+        # Old code:
+        # survival_at_t = (surv_array + (T_neighs > t)).min(axis=1)
+
+        mask = np.zeros_like(T_neighs)
+        mask[T_neighs > t] = np.inf
+        survival_at_t = (surv_array + mask).min(axis=1)
+        # survival_at_t = np.where(~np.isfinite(survival_at_t), survival_at_t, 1.0)
+
         surv_df.append(survival_at_t)
 
     surv_df = pd.DataFrame(surv_df, index=time_bins).T
+    surv_df.where(np.isfinite(surv_df), 1.0, inplace=True)
     return surv_df