Custom scoring on GridSearchCV with fold dependent parameter

Question

The problem

I'm working on a Learning to Rank problem where the norm is to have point evaluated predictions but group evaluated model performance.

More concretely, the estimator outputs a continuous variable (much like a Regressor)

> y = est.predict(X); y
array([71.42857143,  0.        , 71.42857143, ...,  0.        ,
       28.57142857,  0.        ])

But the scoring function requires aggregation by query, that is, grouping predictions, similar to the groups parameter sent to GridSearchCV to respect fold partitioning.

> ltr_score(y_true, y_pred, groups=g)
0.023

The roadblock

So far so good. Things go south when providing the custom scoring function to GridSearchCV, I can't dynamically alter the groups parameter from the scoring function according to the CV folds:

from sklearn.model_selection import GridSearchCV
from sklearn.metrics import make_scorer

ltr_scorer = make_scorer(ltr_score, groups=g)  # Here's the problem, g is fixed
param_grid = {...}

gcv = GridSearchCV(estimator=est, groups=g, param_grid=param_grid, scoring=ltr_scorer)

What is the least hacky way to get around this issue?

One (failed) approach

In a similar question, one comment asked/suggested:

Why cant you just store {the grouping column} locally and utilize it if necessary by indexing with the train test indices provided by the splitter?

To which the OP answered "seems feasible". I thought it was feasible as well, but could not make it work. Apparently, GridSearchCV will first consume all cross-validation split index and only then perform the splits, fits, preds and scorings. This means that I can't (seemingly) try to guess at scoring time the original indexes that created the current split subselection.

For the sake of completeness, my code:

class QuerySplitScorer:
    def __init__(self, X, y, groups):
        self._X = np.array(X)
        self._y = np.array(y)
        self._groups = np.array(groups)
        self._splits = None
        self._current_split = None

    def __iter__(self):
        self._splits = iter(GroupShuffleSplit().split(self._X, self._y, self._groups))
        return self

    def __next__(self):
        self._current_split = next(self._splits)
        return self._current_split

    def get_scorer(self):
        def scorer(y_true, y_pred):
            _, test_idx = self._current_split
            return _score(
                y_true=y_true,
                y_pred=y_pred,
                groups=self._groups[test_idx]
            )

Usage:

qss = QuerySplitScorer(X, y_true, g)
gcv = GridSearchCV(estimator=est, cv=qss, scoring=qss.get_scorer(), param_grid=param_grid, verbose=1)
gcv.fit(X, y_true)

It won't work, self._current_split is fixed at the last generated split.

Answer 1

As I understand scoring values are pairs (value,group), but estimator should not work with group. Let cut them in a wrapper but leave them to scorer.

Simple estimator wrapper (may need some polishing to full compliance)

from sklearn.base import BaseEstimator, ClassifierMixin, TransformerMixin, clone
from sklearn.linear_model import LogisticRegression
from sklearn.utils.estimator_checks import check_estimator
#from sklearn.utils.validation import check_X_y, check_array, check_is_fitted
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import make_scorer

class CutEstimator(BaseEstimator):

    def __init__(self, base_estimator):
        self.base_estimator = base_estimator

    def fit(self, X, y):
        self._base_estimator = clone(self.base_estimator)
        self._base_estimator.fit(X,y[:,0].ravel())
        return self

    def predict(self, X):
        return  self._base_estimator.predict(X)

#check_estimator(CutEstimator(LogisticRegression()))

Then we can use it

def my_score(y, y_pred):

    return np.sum(y[:,1])


pagam_grid = {'base_estimator__C':[0.2,0.5]}

X=np.random.randn(30,3)
y=np.random.randint(3,size=(X.shape[0],1))
g=np.ones_like(y)

gs = GridSearchCV(CutEstimator(LogisticRegression()),pagam_grid,cv=3,
             scoring=make_scorer(my_score), return_train_score=True
            ).fit(X,np.hstack((y,g)))

print (gs.cv_results_['mean_test_score']) #10 as 30/3
print (gs.cv_results_['mean_train_score']) # 20 as 30 -30/3

Output:

 [ 10.  10.]
 [ 20.  20.]

Update 1: Hackers way but no change in estimator:

pagam_grid = {'C':[0.2,0.5]}
X=np.random.randn(30,3)
y=np.random.randint(3,size=(X.shape[0]))
g=np.random.randint(3,size=(X.shape[0]))
cv = GroupShuffleSplit (3,random_state=100)
groups_info = {}
for a,b in cv.split(X, y, g):
    groups_info[hash(y[b].tobytes())] =g[b]
    groups_info[hash(y[a].tobytes())] =g[a]

def my_score(y, y_pred):
    global groups_info
    g = groups_info[hash(y.tobytes())]
    return np.sum(g)

gs = GridSearchCV(LogisticRegression(),pagam_grid,cv=cv, 
             scoring=make_scorer(my_score), return_train_score=True,
            ).fit(X,y,groups = g)
print (gs.cv_results_['mean_test_score']) 
print (gs.cv_results_['mean_train_score'])