lecrapaud 0.8.1__tar.gz → 0.8.3__tar.gz

{lecrapaud-0.8.1 → lecrapaud-0.8.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: lecrapaud
-Version: 0.8.1
+Version: 0.8.3
 Summary: Framework for machine and deep learning, with regression, classification and time series analysis
 License: Apache License
 Author: Pierre H. Gallet

{lecrapaud-0.8.1 → lecrapaud-0.8.3}/lecrapaud/feature_selection.py

@@ -135,10 +135,9 @@ class FeatureSelectionEngine:
         logger.info(f"Starting feature selection for TARGET_{target_number}...")
         clean_directory(self.fs_dir_target)
 
-        # Let's start by removing extremly correlated features
-        # This is needed to reduce nb of feature but also for methods such as anova or chi2 that requires independent features
-        # TODO: we could also remove low variance features
-        self.X = self.remove_constant_columns()
+        # Let's start by removing very low variance feature and extremly correlated features
+        # This is needed to reduce nb of feature but also for methods such as anova or chi2 that requires independent, non constant, non full 0 features
+        self.X = self.remove_low_variance_columns()
         features_uncorrelated, features_correlated = self.remove_correlated_features(
             90, vizualize=False
         )
@@ -353,34 +352,31 @@ class FeatureSelectionEngine:
     # Remove correlation
     # ------------------
 
-    def remove_constant_columns(
-        self,
-        threshold: float = 0.99,
-    ) -> pd.DataFrame:
+    def remove_low_variance_columns(self, threshold: float = 1e-10) -> pd.DataFrame:
         """
-            Removes constant or almost constant columns from a DataFrame.
+        Removes columns with very low variance (including constant columns).
 
         Parameters:
-            threshold (float): Max proportion for a single value (default 0.99 = 99%).
+            threshold (float): Minimum variance required to keep a column.
+                            Default is 1e-10 to eliminate near-constant features.
 
         Returns:
-            pd.DataFrame: Cleaned DataFrame.
+            pd.DataFrame: Cleaned DataFrame without low-variance columns.
         """
+        X = self.X
 
-        to_drop = []
-
-        for col in self.data.columns:
-            value_counts = self.data[col].value_counts(dropna=False, normalize=True)
-            if value_counts.empty:
-                to_drop.append(col)
-            elif value_counts.iloc[0] >= threshold:
-                to_drop.append(col)
+        low_var_cols = [
+            col
+            for col in X.columns
+            if pd.api.types.is_numeric_dtype(X[col])
+            and np.nanvar(X[col].values) < threshold
+        ]
 
-        if to_drop:
-            logger.info(f"🔍 Removed {len(to_drop)} constant/almost constant columns:")
-            logger.info(to_drop)
+        if low_var_cols:
+            logger.info(f"🧹 Removed {len(low_var_cols)} low-variance columns:")
+            logger.info(low_var_cols)
 
-        return self.data.drop(columns=to_drop, errors="ignore")
+        return X.drop(columns=low_var_cols, errors="ignore")
 
     def remove_correlated_features(self, corr_threshold: int, vizualize: bool = False):
         X = self.X

{lecrapaud-0.8.1 → lecrapaud-0.8.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "lecrapaud"
-version = "0.8.1"
+version = "0.8.3"
 description = "Framework for machine and deep learning, with regression, classification and time series analysis"
 authors = [
     {name = "Pierre H. Gallet"}