Scikit learn ordinal classifier according to https://link.springer.com/chapter/10.1007/3-540-44795-4_13

ordinal_classifier.py

# adapted from https://towardsdatascience.com/simple-trick-to-train-an-ordinal-regression-with-any-classifier-6911183d2a3c
class OrdinalClassifier:
    def __init__(self, clf):
        self.clf = clf
        self.clfs = {}

    def fit(self, X, y):
        self.classes_ = np.sort(np.unique(y))
        if self.classes_.shape[0] > 2:
            for level in self.classes_[: (self.classes_.shape[0] - 1)]:
                # for each k - 1 ordinal value we fit a binary classification problem
                binary_y = (y > level).astype(np.uint8)
                clf = clone(self.clf)
                clf.fit(X, binary_y)
                self.clfs[level] = clf

    def predict_proba(self, X):
        clfs_predict = {k: self.clfs[k].predict_proba(X) for k in self.clfs}
        predicted = []
        for previous_level, level in zip([None, *self.classes_], self.classes_):
            if previous_level is None:
                # V1 = 1 - Pr(y > V1)
                predicted.append(1 - clfs_predict[level][:, 1])
            elif level in clfs_predict:
                # Vi = Pr(y > Vi-1) - Pr(y > Vi)
                predicted.append(clfs_predict[previous_level][:, 1] - clfs_predict[level][:, 1])
            else:
                # Vk = Pr(y > Vk-1)
                predicted.append(clfs_predict[previous_level][:, 1])
        decision = np.vstack(predicted).T
        larger_zero = np.where(decision < 0, 0, decision)
        probs = np.where(larger_zero > 1, 1, larger_zero)
        normalized = probs / probs.sum(axis=1).reshape((-1, 1))
        return normalized

    def predict(self, X):
        return self.classes_[np.argmax(self.predict_proba(X), axis=1)]