recommenders.py 文件源码

def recommend(self, ids: Sequence[int],
                  predictions: numpy.ndarray,
                  n: int=1, diversity: float=0.5) -> Sequence[int]:
        """Recommends an instance to label.

        Notes
        -----
        Assumes predictions are probabilities of positive binary label.

        Parameters
        ----------
        ids
            Sequence of IDs in the unlabelled data pool.
        predictions
            N x 1 x C array of predictions. The ith row must correspond with the
            ith ID in the sequence.
        n
            Number of recommendations to make.
        diversity
            Recommendation diversity in [0, 1].

        Returns
        -------
        Sequence[int]
            IDs of the instances to label.
        """
        if predictions.shape[1] != 1:
            raise ValueError('Uncertainty sampling must have one predictor')

        assert len(ids) == predictions.shape[0]

        # x* = argmin p(y1^ | x) - p(y2^ | x) where yn^ = argmax p(yn | x)
        # (Settles 2009).
        partitioned = numpy.partition(predictions, -2, axis=2)
        most_likely = partitioned[:, 0, -1]
        second_most_likely = partitioned[:, 0, -2]
        assert most_likely.shape == (len(ids),)
        scores = 1 - (most_likely - second_most_likely)

        indices = choose_boltzmann(self._db.read_features(ids), scores, n,
                                   temperature=diversity * 2)
        return [ids[i] for i in indices]


# For safe string-based access to recommender classes.