Combine several classifiers' predictions:
hard_vote(preds): preds is a list of label arrays, one per model. Each sample gets the most common label (ties: the smallest label).soft_vote(probas, weights=None): probas is a list of (n, classes) probability arrays. Average them (weighted if weights is given) and take the argmax.agreement(preds): the fraction of samples where every model agrees, a quick diversity check (ensembles help most when models disagree)