def _ndcg_score(y_true, y_pred, qid, k=None, dcg_func=None, idcg_cache=None):
assert dcg_func is not None
y_true = np.maximum(y_true, 0)
dcg = _dcg_score(y_true, y_pred, qid, k=k, dcg_func=dcg_func)
if idcg_cache:
idcg = []
for a, b in group_offsets(qid):
if qid[a] not in idcg_cache:
idcg_cache[qid[a]] = dcg_func(np.sort(y_true[a:b]),
np.arange(0, b - a),
k=k)
idcg.append(idcg_cache[qid[a]])
else:
idcg = [
dcg_func(np.sort(y_true[a:b]), np.arange(0, b - a), k=k)
for a, b in group_offsets(qid)
]
idcg = np.array(idcg)
assert (dcg <= idcg).all()
idcg[idcg == 0] = 1
return dcg / idcg
def MAPScorer(y_true, y_pred, qid):
return np.mean(np.array([_map_score(y_true[a:b], y_pred[a:b]) for a, b in group_offsets(qid)]))
def r_precision(y_true, y_pred, qid):
return np.mean(np.array([rp(y_true[a:b], y_pred[a:b]) for a, b in group_offsets(qid)]))
def mean_reciprocal_rank(y_true, y_pred, qid):
# MAP (Mean Average Precision)
return np.mean(np.array([mrr(y_true[a:b], y_pred[a:b]) for a, b in group_offsets(qid)]))
def _dcg_score(y_true, y_pred, qid, k=None, dcg_func=None):
assert dcg_func is not None
y_true = np.maximum(y_true, 0)
return np.array([
dcg_func(y_true[a:b], y_pred[a:b], k=k) for a, b in group_offsets(qid)
])
def p_score(y_true, y_pred, qid, k=None):
return np.array([
_p_score(y_true[a:b], y_pred[a:b], k=k) for a, b in group_offsets(qid)
])