Logreg future warning

docs/source/notes/cli.rst

@@ -79,7 +79,9 @@ Full CLI options (check out with ``geneplexus --help``)
 	                        123})
 	  -lk , --logreg_kwargs
 	                        Set the logistic regression keyword arguments in fit(). (default:
-	                        {'max_iter': 10000, 'solver': 'lbfgs', 'penalty': 'l2', 'C': 1.0})
+	                        {'max_iter': 10000, 'solver': 'lbfgs', 'C': 1.0, 'l1_ratio': 0.0} on
+	                        scikit-learn>=1.8, otherwise {'max_iter': 10000, 'solver': 'lbfgs',
+	                        'penalty': 'l2', 'C': 1.0})
 	  -s, --scale           When added, will set scale to True in fit(). See docs for more info of
 	                        when this is good to do. (default: False)
 	  -mnp , --min_num_pos

geneplexus/_config/config.py

@@ -1,6 +1,7 @@
 """Global variables used by the GenePlexus library."""
 import os.path as osp
 import pathlib
+import re
 from typing import Any
 from typing import Dict
 from typing import List
@@ -10,6 +11,7 @@
 from typing import Union
 
 import numpy as np
+import sklearn
 
 MAX_RETRY = 10  # maximum number of retries for downloading
 
@@ -118,9 +120,20 @@
 DEFAULT_LOGREG_KWARGS: Dict[str, Any] = {
     "max_iter": 10000,
     "solver": "lbfgs",
-    "penalty": "l2",
     "C": 1.0,
 }
+# In scikit-learn>=1.8, ``penalty`` is deprecated in favor of ``l1_ratio``/``C``
+# (l1_ratio=0 is equivalent to the old penalty="l2" default), and passing
+# ``penalty`` explicitly now raises a FutureWarning even when set to "l2".
+# Older versions don't accept l1_ratio unless penalty="elasticnet", so pick
+# whichever explicit form matches the installed version instead of relying on
+# either version's default.
+_SKLEARN_VERSION_MATCH = re.match(r"(\d+)\.(\d+)", sklearn.__version__)
+_SKLEARN_VERSION = tuple(map(int, _SKLEARN_VERSION_MATCH.groups())) if _SKLEARN_VERSION_MATCH else (0, 0)
+if _SKLEARN_VERSION >= (1, 8):
+    DEFAULT_LOGREG_KWARGS["l1_ratio"] = 0.0
+else:
+    DEFAULT_LOGREG_KWARGS["penalty"] = "l2"
 
 __all__ = [
     "URL_DICT",

geneplexus/cli.py

@@ -225,7 +225,10 @@ def parse_args() -> argparse.Namespace:
         default=config.DEFAULT_LOGREG_KWARGS,
         metavar="",
         type=json.loads,
-        help="Set the logistic regression keyword arguments in fit().",
+        help=(
+            "Set the logistic regression keyword arguments in fit(). The default shown "
+            "depends on the installed scikit-learn version (penalty vs l1_ratio)."
+        ),
     )
 
     parser.add_argument(

geneplexus/geneplexus.py

@@ -630,7 +630,10 @@ def fit(
 
         Args:
             logreg_kwargs: Scikit-learn logistic regression settings (see
-                :class:`~sklearn.linear_model.LogisticRegression`).
+                :class:`~sklearn.linear_model.LogisticRegression`). The default
+                uses ``l1_ratio=0.0`` on scikit-learn>=1.8 and ``penalty="l2"``
+                on older versions (equivalent settings), since ``penalty`` is
+                deprecated in scikit-learn>=1.8.
             scale: Whether to scale the data when doing model training and prediction. It is
                 not recommended to set to ``True`` unless using custom data.
             min_num_pos: Minimum number of positives required for the model