def confidence_calibration_check(round_idx, respondent_idx, d_metric_dict,
n_neighbors):
_, X_test, _, y_test, _ = load_dataset(onehot=True)
X_test = preprocessing_data_with_unit_var(X_test)
d_metric = d_metric_dict['R%d_W%d' % (respondent_idx, round_idx)]
weighted_y_test = pd.DataFrame(
weighted_prediction_round(round_idx)[respondent_idx - 1].values,
index=y_test.index,
columns=['weighted_y_test'])
weighted_y_test['y_test'] = y_test
weighted_y_test['neighbor reci prop'] = y_test
for idx in X_test.index:
diff = X_test.values - X_test.loc[idx].values
X_test['distance'] = np.dot(diff @ d_metric, diff.T).diagonal()
max_d = X_test['distance'].sort_values(
ascending=True).values[n_neighbors]
neighbor_list = [
my_idx
for my_idx in X_test.loc[X_test['distance'] <= max_d].index.values
if my_idx != idx
]
reci = 0
for neighbor in neighbor_list:
reci += weighted_y_test.loc[neighbor, 'y_test']
weighted_y_test.loc[neighbor,
'neighbor reci prop'] = reci / n_neighbors
X_test.drop(['distance'], axis=1, inplace=True)
return weighted_y_test
def similarity_fig_from_metric_by_nca_part(round_idx, respondent_idx,
n_neighbors, d_metric_dict):
"""
similar as nca, only here the neighborhood is defined as the minimal hypersphere
that contains the n_neighbors similar points of x_i
:param round_idx: integer, ranging from 1 to 10
:param respondent_idx: id of respondents, ranging from the 1 to 35
:param n_neighbors: the number of similar points to define the neighborhood
:param d_metric_dict: dict to save learned distance metric, indexed by the
respondent_id and the round_idx
:return: a similarity connection graph and the similarity edge list
"""
_, X_test, _, y_test, _ = load_dataset(onehot=True)
X_test = preprocessing_data_with_unit_var(X_test)
d_metric = d_metric_dict['R%d_W%d' % (respondent_idx, round_idx)]
# weighted_y_test = pd.DataFrame(weighted_prediction(round_idx)[respondent_idx - 1].values,
# index=y_test.index)
weighted_y_test = pd.DataFrame(weighted_prediction_quantile(
weighted_prediction_round(round_idx))[respondent_idx - 1].values,
index=y_test.index)
similarity_edges = []
for idx in X_test.index:
diff = X_test.values - X_test.loc[idx].values
X_test['distance'] = np.dot(diff @ d_metric, diff.T).diagonal()
same_class = X_test.loc[(
weighted_y_test == weighted_y_test.loc[idx]).values.ravel()]
max_d = same_class['distance'].sort_values(ascending=True).values[min(
len(same_class) - 1, n_neighbors)]
#print('idx: %5d'%idx, ' max_d: %2.2f.'%max_d)
neighbor_list = [
my_idx
for my_idx in X_test.loc[X_test['distance'] <= max_d].index.values
if my_idx != idx
]
for item in neighbor_list:
similarity_edges.append((idx, item))
X_test.drop(['distance'], axis=1, inplace=True)
fig, edges = similarity_fig_from_weighted_prediction(
round_idx=round_idx,
respondent_idx=respondent_idx,
similarity_edge_list=similarity_edges)
return fig, edges
def similarity_fig_from_metric_by_nca(round_idx, respondent_idx,
d_metric_dict):
"""
nca -> neighborhood component analysis
First is to define the neighborhood of data point x_i in the transformed space,
it is the minimal hypersphere which centers at x_i, and its radius is equal to
the farthest point which is similar with x_i. This function is to build similarity
edges between x_i and all the data points within its neighborhood.
:param round_idx: integer, ranging from 1 to 10
:param respondent_idx: id of respondents, ranging from the 1 to 35
:param d_metric_dict: dict to save learned distance metric, indexed by the
respondent_id and the round_idx
:return: a similarity connection graph and the similarity edge list
"""
_, X_test, _, y_test, _ = load_dataset(onehot=True)
X_test = preprocessing_data_with_unit_var(X_test)
d_metric = d_metric_dict['R%d_W%d' % (respondent_idx, round_idx)]
weighted_y_test = pd.DataFrame(weighted_prediction_quantile(
weighted_prediction_round(round_idx))[respondent_idx - 1].values,
index=y_test.index)
similarity_edges = []
for idx in X_test.index:
diff = X_test.values - X_test.loc[idx].values
X_test['distance'] = np.dot(diff @ d_metric, diff.T).diagonal()
max_d = X_test.loc[(weighted_y_test == weighted_y_test.loc[idx]
).values.ravel()]['distance'].max()
#print('idx: %5d' % idx, ' max_d: %2.2f.' % max_d)
neighbor_list = [
my_idx
for my_idx in X_test.loc[X_test['distance'] <= max_d].index.values
if my_idx != idx
]
for item in neighbor_list:
similarity_edges.append((idx, item))
X_test.drop(['distance'], axis=1, inplace=True)
fig, edges = similarity_fig_from_weighted_prediction(
round_idx=round_idx,
respondent_idx=respondent_idx,
similarity_edge_list=similarity_edges)
return fig, edges
def cal_distance_with_df(d_metric_dict):
_, X_test, _, y_test, _ = load_dataset(onehot=True)
X_test = preprocessing_data_with_unit_var(X_test)
distance_dict = dict()
for round_idx in range(1, 11):
for respondent_idx in range(1, 36):
distance_df = pd.DataFrame(np.zeros((len(X_test), len(X_test))),
index=X_test.index,
columns=X_test.index)
d_metric = d_metric_dict['R%d_W%d' % (respondent_idx, round_idx)]
for idx in X_test.index:
diff = X_test.values - X_test.loc[idx].values
distance_df[idx] = np.dot(diff @ d_metric, diff.T).diagonal()
distance_dict["R%d_W%d" %
(respondent_idx, round_idx)] = distance_df
print(respondent_idx, round_idx)
np.save(r'../new_data/lmnn_distance.npy', distance_dict)
return None
def similarity_edges_from_metric_by_knn(round_idx, respondent_idx, n_neighbors,
d_metric_dict):
_, X_test, _, y_test, _ = load_dataset(onehot=True)
X_test = preprocessing_data_with_unit_var(X_test)
d_metric = d_metric_dict['R%d_W%d' % (respondent_idx, round_idx)]
similarity_edges = {}
for idx in X_test.index:
diff = X_test.values - X_test.loc[idx].values
X_test['distance'] = np.dot(diff @ d_metric, diff.T).diagonal()
max_d = X_test['distance'].sort_values(
ascending=True).values[n_neighbors]
neighbor_list = [
my_idx
for my_idx in X_test.loc[X_test['distance'] <= max_d].index.values
if my_idx != idx
]
similarity_edges[idx] = neighbor_list
X_test.drop(['distance'], axis=1, inplace=True)
return similarity_edges
def similarity_fig_from_metric_by_thres(round_idx, respondent_idx,
d_metric_dict, d_thres):
"""
Basically to build the similarity edges between node_i and node_j
if there d^2(node_i, node_j) within the transformed feature space is
less than the threshold value, d_thres.
:param round_idx: integer, ranging from 1 to 10
:param respondent_idx: id of respondents, ranging from the 1 to 35
:param d_metric_dict: dict to save learned distance metric, indexed by the
respondent_id and the round_idx
:param d_thres: threshold value to determine whether the similarity
edges should be built
:return: a similarity connection graph and the similarity edge list
"""
_, X_test, _, y_test, _ = load_dataset(onehot=True)
X_test = preprocessing_data_with_unit_var(X_test)
d_metric = d_metric_dict['R%d_W%d' % (respondent_idx, round_idx)]
similarity_edges = []
for idx in X_test.index:
diff = X_test.values - X_test.loc[idx].values
X_test['distance'] = np.dot(diff @ d_metric, diff.T).diagonal()
neighbor_list = [
my_idx for my_idx in X_test.loc[
X_test['distance'] <= d_thres].index.values if my_idx != idx
]
for item in neighbor_list:
similarity_edges.append((idx, item))
X_test.drop(['distance'], axis=1, inplace=True)
fig, edges = similarity_fig_from_weighted_prediction(
round_idx=round_idx,
respondent_idx=respondent_idx,
similarity_edge_list=similarity_edges)
return fig, edges