def _calc_score(predict_y, true_y, nlabels, with_cfmat=False):
"""
Calculate various scoring of the prediction.
:param predict_y: predicted value (enumerated)
:param true_y: actual value (enumerated)
:param nlabels: number of labels (true_y might not have all labels)
:param with_cfmat: if True will return confusiion matrix
:return: score( percentage correct)
label_hits (ndarray of hit count for each label. Labels not present in true_y will have count of NaN)
label_misses similar to label_hits but count the misses
cf_matrix (optional) the confusion matrix (nlabels x nlabels)
"""
hits = (predict_y == true_y).astype(np.int)
misses = 1 - hits
score = hits.sum() / len(true_y)
label_hits = np.full((nlabels, ), np.nan)
label_misses = np.full((nlabels, ), np.nan)
unique_test_labels = np.unique(true_y)
_label_hits = accum(true_y, hits, func=np.sum, dtype=np.int)
_label_misses = accum(true_y, misses, func=np.sum, dtype=np.int)
label_hits[unique_test_labels] = _label_hits[unique_test_labels]
label_misses[unique_test_labels] = _label_misses[unique_test_labels]
if with_cfmat:
cf_matrix = confusion_matrix(true_y, predict_y)
return score, label_hits, label_misses, cf_matrix
else:
return score, label_hits, label_misses
def sorted_labels_by_prevalence_then_distance(neighbour_labels,
neighbour_distances,
distance_func):
# labels = np.array(['a', 'a', 'b', 'c', 'b', 'a', 'c', 'd'])
# dists = np.array([0.1, 0.1, 0.3, 0.3, 0.5, 0.2, 0.4, 0.6])
# indices = np.array([0, 1, 2, 3, 4, 5, 6, 7])
sorted_indices = np.argsort(neighbour_distances)
closests = sorted_indices[0:7]
closests_labels = neighbour_labels[closests]
unique_labels, indices = np.unique(closests_labels, return_inverse=True)
prevalences = np.bincount(indices)
closests_distances = neighbour_distances[closests]
mean_distance_by_labels = accum(indices,
closests_distances,
func=distance_func,
dtype=np.float)
# Now, lexsort sorts everything ascendingly. We want the prevalences to be sorted descendingly and distances
# ascendingly. So we must negate prevalences
sorted_indices_by_prevalence_then_mean_distance = np.lexsort(
(0 - prevalences, mean_distance_by_labels))
sorted_labels = unique_labels[
sorted_indices_by_prevalence_then_mean_distance]
return sorted_labels
def _calc_score(predict_y, test_y, nlabels):
hits = (predict_y == test_y).astype(np.int)
misses = 1 - hits
score = hits.sum() / len(test_y)
label_hits = np.full((nlabels, ), np.nan)
label_misses = np.full((nlabels, ), np.nan)
unique_test_labels = np.unique(test_y)
_label_hits = accum(test_y, hits, func=np.sum, dtype=np.int)
_label_misses = accum(test_y, misses, func=np.sum, dtype=np.int)
label_hits[unique_test_labels] = _label_hits[unique_test_labels]
label_misses[unique_test_labels] = _label_misses[unique_test_labels]
return score, label_hits, label_misses
def k_nearest(distmat, train_labels, nlabels, k, map_order):
# Leave one out K nearest neighbour
# Find the nearest 5 neighbours & find out which label dominates
element_count = distmat.shape[0]
actual_labels = []
predicted_labels = []
for j in range(element_count):
distances_from_j = distmat[j, :]
sorted_indices = np.argsort(distances_from_j)
closests_indices = sorted_indices[1:k + 1]
closests_labels = train_labels[closests_indices]
closests_distances = distances_from_j[closests_indices]
predicted_label = sorted_labels_by_prevalence_then_distance(
closests_labels, closests_distances, np.mean)
predicted_labels.append(predicted_label)
actual_labels.append([train_labels[j]])
label_prediction_map_score = mapk(actual_labels, predicted_labels,
map_order)
hits = np.array(
[1 if a in p else 0 for a, p in zip(actual_labels, predicted_labels)],
dtype=np.int)
misses = np.array(
[0 if a in p else 1 for a, p in zip(actual_labels, predicted_labels)],
dtype=np.int)
label_hits = np.full((nlabels, ), np.nan)
label_misses = np.full((nlabels, ), np.nan)
unique_test_labels = np.unique(train_labels)
_label_hits = accum(train_labels, hits, func=np.sum, dtype=np.int)
_label_misses = accum(train_labels, misses, func=np.sum, dtype=np.int)
label_hits[unique_test_labels] = _label_hits[unique_test_labels]
label_misses[unique_test_labels] = _label_misses[unique_test_labels]
return label_prediction_map_score, label_hits, label_misses