Consider multiplying the quadratic biases in SelectFromQuadraticModel.correlation_cqm() by 1-alpha.

[arcondello]

Currently, it is possible to construct a dataset where, even with alpha=1, a fixed or random column is chosen because of the penalties in the quadratic term. This contradicts the docstring which claims

            alpha:
                Hyperparameter between 0 and 1 that controls the relative weight of
                the relevance and redundancy terms.
                ``alpha=0`` places no weight on the quality of the features,
                therefore the features will be selected as to minimize the
                redundancy without any consideration to quality.
                ``alpha=1`` places the maximum weight on the quality of the features,
                and therefore will be equivalent to using
                :class:`sklearn.feature_selection.SelectKBest`.

One solution is to multiply the quadratic/redundany terms by 1-alpha to ensure that they are zeroed when alpha=1.

For instance we could replace

dwave-scikit-learn-plugin/dwave/plugins/sklearn/transformers.py

Lines 210 to 212 in 7db7d03

	# our objective
	# we multiply by 2 because the matrix is symmetric
	np.fill_diagonal(correlations, correlations[:, -1] * (-2 * alpha * num_features))

with

            diag = np.array(correlations[:, -1] * (-2 * alpha * num_features), copy=True)

            correlations *= (1-alpha)

            # our objective
            # we multiply by 2 because the matrix is symmetric
            np.fill_diagonal(correlations, diag)