def jaccard_similarity(pred_0, pred_1, y=None):
"""Jaccard similary b/w two different predictions matrices
Args:
pred_0: csr_matrix
prediction for algorithm 0
pred_1: csr_matrix
prediction for algorithm 1
y: csr_matrix or None
true labels
"""
def _correct_only(pred, y):
pred = pred.multiply(y)
pred.eliminate_zeros()
return pred
def _safe_divide(a, b):
with np.errstate(divide='ignore', invalid='ignore'):
out = np.true_divide(a, b)
out[out == np.inf] = 0
return np.nan_to_num(out)
if y is not None:
pred_0 = _correct_only(pred_0, y)
pred_1 = _correct_only(pred_1, y)
pred_0, pred_1 = binarize(pred_0), binarize(pred_1)
intersection = np.array(pred_0.multiply(pred_1).sum(axis=1)).ravel()
union = np.array(binarize(pred_0 + pred_1).sum(axis=1)).ravel()
return np.mean(_safe_divide(intersection, union))
def recall(predicted_labels, true_labels, k=5):
"""Compute recall@k
Args:
predicted_labels: csr_matrix
predicted labels
true_labels: csr_matrix
true_labels
k: int, default=5
keep only top-k predictions
"""
predicted_labels = retain_topk(predicted_labels, k)
denom = np.sum(true_labels, axis=1)
rc = binarize(true_labels.multiply(predicted_labels))
rc = np.sum(rc, axis=1) / (denom + 1e-5)
return np.mean(rc) * 100