PLSRegresson supports multi-target regression, hence the predicition is 2d-array.

As a proposed solution, I suggest creating a custom subclass of PLSRegression, which we can call PLS1D. This subclass overrides the predict method of PLSRegression to ensure it returns a 1-dimensional array, making it compatible with the VotingRegressor. This will effectively handle the case where n_components=1.

Here is the relevant code for the proposed PLS1D subclass:

import numpy as np
from sklearn.linear_model import LinearRegression
from sklearn.cross_decomposition import PLSRegression
from sklearn.ensemble import VotingRegressor



class PLS1D(PLSRegression):
    def __init__(self, n_components=1, scale=True, max_iter=500, tol=1e-06, copy=True):
        if n_components > 1:
            raise ValueError("PLS1D can only handle n_components=1")
        super().__init__(n_components=n_components, scale=scale, max_iter=max_iter, tol=tol, copy=copy)
    
    def predict(self, X):
        y_pred = super().predict(X)
        return y_pred.ravel()


r1 = LinearRegression()
r2 = PLS1D()

X = np.array([[1, 1], [2, 4], [3, 9], [4, 16], [5, 25], [6, 36]])
y = np.array([2, 6, 12, 20, 30, 42])
er = VotingRegressor([('lr', r1), ('plsr', r2)])
print(er.fit(X, y).predict(X))

r1.fit(X, y).predict(X).shape
r2.fit(X, y).predict(X).shape

Output
[ 1.21007429 6.09332284 12.55111269 20.58344384 30.1903163 41.37173005]

I think that we might want PLSRegression to behave consistently with LinearRegression and Ridge, that is, when y is provided as shape (n_samples, 1) at fit, then predict should output that same shape, but when y is passed as shape (n_samples,) then predict would automatically ravel its output.

I think other multi-output regressors such as RandomForestRegressor and other tree-based regressors should also do the same.

We might also accept a PR to make VotingRegressor regressor do such raveling automatically when all inputs have a mix of (n_samples,) and (n_samples, 1) shapes.

I would like to make a PR for the first suggestion (i.e., automatically ravel the prediction when Y.ndim == 1 when fitting). I often needed to ravel manually when using scikit-learn so I think this is desirable, but I do think this is a breaking change.

/take