evaluation.py 文件源码

def get_ranking_scores(matched_predictions, feedback_data, switch_positive, alternative=True):
    users_num, topk, holdout = matched_predictions.shape
    ideal_scores_idx = np.argsort(feedback_data, axis=1)[:, ::-1] #returns column index only
    ideal_scores_idx = np.ravel_multi_index((np.arange(feedback_data.shape[0])[:, None], ideal_scores_idx), dims=feedback_data.shape)

    where = np.ma.where if np.ma.is_masked(feedback_data) else np.where
    is_positive = feedback_data >= switch_positive
    positive_feedback = where(is_positive, feedback_data, 0)
    negative_feedback = where(~is_positive, -feedback_data, 0)

    relevance_scores_pos = (matched_predictions * positive_feedback[:, None, :]).sum(axis=2)
    relevance_scores_neg = (matched_predictions * negative_feedback[:, None, :]).sum(axis=2)
    ideal_scores_pos = positive_feedback.ravel()[ideal_scores_idx]
    ideal_scores_neg = negative_feedback.ravel()[ideal_scores_idx]

    discount_num = max(holdout, topk)
    if alternative:
        discount = np.log2(np.arange(2, discount_num+2))
        relevance_scores_pos = 2**relevance_scores_pos - 1
        relevance_scores_neg = 2**relevance_scores_neg - 1
        ideal_scores_pos = 2**ideal_scores_pos - 1
        ideal_scores_neg = 2**ideal_scores_neg - 1
    else:
        discount = np.hstack([1, np.log(np.arange(2, discount_num+1))])

    dcg = (relevance_scores_pos / discount[:topk]).sum(axis=1)
    dcl = (relevance_scores_neg / -discount[:topk]).sum(axis=1)
    idcg = (ideal_scores_pos / discount[:holdout]).sum(axis=1)
    idcl = (ideal_scores_neg / -discount[:holdout]).sum(axis=1)

    with np.errstate(invalid='ignore'):
        ndcg = unmask(np.nansum(dcg / idcg) / users_num)
        ndcl = unmask(np.nansum(dcl / idcl) / users_num)

    ranking_score = namedtuple('Ranking', ['nDCG', 'nDCL'])._make([ndcg, ndcl])
    return ranking_score