feat: split .predict method in .predict and .predict_set, add "QUESTI…

…ON" tag in comment to identify important points to figure out, add few corrections and comments

public_api_v1_regression.py

@@ -1,9 +1,8 @@
-from typing import Optional, Union, Self, Iterable, Tuple, List
+from typing import Optional, Union, Self, Tuple, List
 
 import numpy as np
 from sklearn.linear_model import LinearRegression, QuantileRegressor
 
-from mapie.regression import MapieQuantileRegressor
 from numpy.typing import ArrayLike, NDArray
 from sklearn.base import RegressorMixin
 from sklearn.model_selection import BaseCrossValidator
@@ -14,9 +13,9 @@
 class NaiveConformalRegressor:
  def __init__(
  self,
- estimator: RegressorMixin = LinearRegression, # Improved 'None' default
- conformity_score: BaseRegressionScore = AbsoluteConformityScore, # Should we set this default?
- alpha: Union[float, List[float]] = 0.1, # Should we set this default? I think an array is OK (already implemented, and avoid developing a less user-friendly reset_alpha method)
+ estimator: RegressorMixin = LinearRegression(), # Improved 'None' default
+ conformity_score: Union[str, BaseRegressionScore] = "absolute", # Add string option
+ alpha: Union[float, List[float]] = 0.1, # QUESTION: Should we set this default, or should we keep None? I think an array is OK (already implemented, and avoid developing a less user-friendly reset_alpha method)
  n_jobs: Optional[int] = None,
  verbose: int = 0,
  random_state: Optional[Union[int, np.random.RandomState]] = None,
@@ -28,39 +27,51 @@ def fit(
  X: ArrayLike,
  y: ArrayLike,
  # sample_weight: Optional[ArrayLike] = None, -> in fit_params
- fit_params: Optional[dict] = None, # -> In __init__ ?
- predict_params: Optional[dict] = None, # -> In __init__ ?
+ fit_params: Optional[dict] = None, # Ex for LGBMClassifier: {'categorical_feature': 'auto'}
+ predict_params: Optional[dict] = None,
  ) -> Self:
  pass
 
- def predict(
+ def predict_set(
  self,
  X: ArrayLike,
- optimize_beta: bool = False, # Don't understand that one
+ optimize_beta: bool = False,
  allow_infinite_bounds: bool = False,
+ # **predict_params -> QUESTION: Is this redundant with predict_params in .fit() ?
+ ) -> NDArray:
+ """
+ Returns
+ -------
+ An array containing the prediction intervals,
+ of shape (n_samples, 2) if alpha is a float,
+ or (n_samples, 2, n_alpha) if alpha is an array of floats
+ """
+ pass
+
+ def predict(
+ self,
+ X: ArrayLike,
  # **predict_params -> Is this redundant with predict_params in .fit() ?
- ) -> Tuple[NDArray, NDArray]:
+ ) -> NDArray:
  """
  Returns
  -------
- Tuple[NDArray, NDArray]:
- - the first element contains the point predictions, with shape (n_samples,)
- - the second element contains the prediction intervals,
- with shape (n_samples, 2) if alpha is a float, or (n_samples, 2, n_alpha) if alpha is an array of floats
+ An array containing the point predictions, with shape (n_samples,)
  """
  pass
 
 
 class SplitConformalRegressor:
  def __init__(
  self,
- estimator: RegressorMixin = LinearRegression, # Improved 'None' default
- conformity_score: BaseRegressionScore = AbsoluteConformityScore, # Should we set this default?
+ estimator: RegressorMixin = LinearRegression(), # Improved 'None' default
+ conformity_score: Union[str, BaseRegressionScore] = "absolute", # Add string option
  alpha: Union[float, List[float]] = 0.1, # See comment in NaiveConformalRegressor
- split_method: str = "simple", # 'simple' (provide test_size in .fit) or 'prefit'. Future API: 'manual' (provide X_calib, Y_calib in predict) and BaseCrossValidator (restricted to splitters only)
+ split_method: str = "simple", # 'simple' (provide test_size in .fit) or 'prefit'. Future API: 'manual' (provide X_calib, Y_calib in fit) and BaseCrossValidator (restricted to splitters only)
  n_jobs: Optional[int] = None,
  verbose: int = 0,
  random_state: Optional[Union[int, np.random.RandomState]] = None
+ # groups -> not used in the current implementation (that is using ShuffleSplit)
  ) -> None:
  pass
 
@@ -72,22 +83,22 @@ def fit(
  test_size: Union[int, float] = 0.1, # Moved from __init__, improved 'None' default. Invalid if split_method != 'simple'
  # Future API: X_calib: Optional[ArrayLike] = None, # Must be None if split_method != 'manual'
  # Future API: y_calib: Optional[ArrayLike] = None, # Must be None if split_method != 'manual'
- fit_params: Optional[dict] = None, # -> In __init__ ?
- predict_params: Optional[dict] = None, # -> In __init__ ?
+ fit_params: Optional[dict] = None,
+ predict_params: Optional[dict] = None,
  ) -> Self:
  pass
 
- # predict signature are the same as NaiveConformalRegressor
+ # predict and predict_set signatures are the same as NaiveConformalRegressor
 
 
 class CrossConformalRegressor:
  def __init__(
  self,
- estimator: RegressorMixin = LinearRegression, # Improved 'None' default
- conformity_score: BaseRegressionScore = AbsoluteConformityScore, # Should we set this default?
+ estimator: RegressorMixin = LinearRegression(), # Improved 'None' default
+ conformity_score: Union[str, BaseRegressionScore] = "absolute", # Add string option
  alpha: Union[float, List[float]] = 0.1, # See comment in NaiveConformalRegressor
  method: str = "plus", # 'base' | 'plus' | 'minmax'
- cross_val: Union[int, BaseCrossValidator] = None, # Improved 'None' default, removed str option, update name. Note that we lose the prefit option, that was I think useless in a cross-validation context
+ cross_val: Union[int, BaseCrossValidator] = 5, # Improved 'None' default, removed str option, update name. Note that we lose the prefit option, that was I think useless in a cross-validation context QUESTION
  # agg_function -> moved to predict method
  n_jobs: Optional[int] = None,
  verbose: int = 0,
@@ -101,27 +112,36 @@ def fit(
  y: ArrayLike,
  # sample_weight: Optional[ArrayLike] = None, -> in fit_params
  # groups: Optional[ArrayLike] = None, -> To specify directly in the cross_val parameter
- fit_params: Optional[dict] = None, # -> In __init__ ?
- predict_params: Optional[dict] = None, # -> In __init__ ?
+ fit_params: Optional[dict] = None,
+ predict_params: Optional[dict] = None,
  ) -> Self:
  pass
 
- def predict(
+ def predict_set(
  self,
  X: ArrayLike,
- # ensemble: bool = False, -> replaced by aggregation_strategy
- aggregation_strategy: Optional[str] = None, # If None, the paper implementation is used
- optimize_beta: bool = False, # Don't understand that one
+ optimize_beta: bool = False,
  allow_infinite_bounds: bool = False,
  # **predict_params -> To remove: redundant with predict_params in .fit()
- ) -> Tuple[NDArray, NDArray]: # See docstring in NaiveConformalRegressor for the return type details
+ ) -> NDArray: # See docstring in NaiveConformalRegressor for the return type details
+ pass
+
+ def predict(
+ self,
+ # ensemble: bool = False, -> removed, see aggregation_method
+ aggregation_method: Optional[str] = None, # None: no aggregation, 'mean', 'median'
+ ) -> NDArray:
  pass
 
 
+class JackknifeAfterBootstrapRegressor:
+ pass # TODO
+
+
 class ConformalizedQuantileRegressor:
  def __init__(
  self,
- estimator: RegressorMixin = QuantileRegressor, # Improved 'None' default
+ estimator: RegressorMixin = QuantileRegressor(), # Improved 'None' default
  alpha: Union[float, List[float]] = 0.1, # See comment in NaiveConformalRegressor
  split_method: str = "simple", # 'simple' (provide test_size in .fit), 'prefit' or 'manual'. Future API: BaseCrossValidator (restricted to splitters only)
  random_state: Optional[Union[int, np.random.RandomState]] = None, # Moved from .fit
@@ -136,21 +156,23 @@ def fit(
  y: ArrayLike,
  # sample_weight: Optional[ArrayLike] = None, -> in fit_params
  # groups: Optional[ArrayLike] = None, -> To specify directly in the cross_val parameter
- # shuffle: Optional[bool] = True, -> To implement in a future version (using the BaseCrossValidator split_method). In that case we would lose that feature in the v1.0.0
+ # shuffle: Optional[bool] = True, -> To implement in a future version (using the BaseCrossValidator split_method). In that case we would lose that feature in the v1.0.0 QUESTION
  # stratify: Optional[ArrayLike] = None, -> same comment as shuffle
  test_size: Union[int, float] = 0.1, # Renamed from 'calib_size'
  X_calib: Optional[ArrayLike] = None, # Must be None if split_method != 'manual'
  y_calib: Optional[ArrayLike] = None, # Must be None if split_method != 'manual'
- fit_params: Optional[dict] = None, # -> In __init__ ?
- predict_params: Optional[dict] = None, # -> In __init__ ?
+ fit_params: Optional[dict] = None,
+ predict_params: Optional[dict] = None,
  ) -> Self:
  pass
 
- def predict(
+ def predict_set(
  self,
  X: ArrayLike,
  optimize_beta: bool = False,
  allow_infinite_bounds: bool = False,
  symmetry: bool = True, # Corrected typing
- ) -> Tuple[NDArray, NDArray]: # See docstring in NaiveConformalRegressor for the return type details
+ ) -> NDArray:
  pass
+
+ # predict signature is the same as NaiveConformalRegressor