-
Notifications
You must be signed in to change notification settings - Fork 9
/
cluster.py
204 lines (180 loc) · 10.8 KB
/
cluster.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
import numpy as np
import yaml
import os
import dill
import pandas as pd
from tqdm import tqdm
from datetime import datetime
from kmodes.kprototypes import KPrototypes
from kmodes.util.dissim import euclidean_dissim, matching_dissim
from tensorflow.keras.models import load_model
from joblib import load
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import silhouette_score
from src.interpretability.lime_explain import predict_and_explain, predict_instance
from src.visualization.visualize import visualize_cluster_explanations, visualize_silhouette_plot
def cluster_clients(k=None, save_centroids=True, save_clusters=True, explain_centroids=True):
'''
Runs k-prototype clustering algorithm on preprocessed dataset
:param k: Desired number of clusters
:param save_centroids: Boolean indicating whether to save cluster centroids
:param save_clusters: Boolean indicating whether to save client cluster assignments
:param explain_centroids: Boolean indicating whether to compute LIME explanations for cluster centroids
:return: A KPrototypes object that describes the best clustering of all the runs
'''
cfg = yaml.full_load(open(os.getcwd() + "/config.yml", 'r'))
# Load preprocessed client data
try:
df = pd.read_csv(cfg['PATHS']['PROCESSED_DATA'])
except FileNotFoundError:
print("No file found at " + cfg['PATHS']['PROCESSED_DATA'] + ". Run preprocessing script before running this script.")
return
client_ids = df.pop('ClientID').tolist()
if cfg['TRAIN']['DATASET_TYPE'] == 'static_and_dynamic':
dates = df.pop('Date').tolist()
df.drop('GroundTruth', axis=1, inplace=True)
X = np.array(df)
# Load feature info
try:
data_info = yaml.full_load(open(cfg['PATHS']['DATA_INFO'], 'r'))
except FileNotFoundError:
print("No file found at " + cfg['PATHS']['DATA_INFO'] + ". Run preprocessing script before running this script.")
return
# Get list of categorical feature indices
noncat_feat_idxs = [df.columns.get_loc(c) for c in data_info['NON_CAT_FEATURES'] if c in df]
cat_feat_idxs = [i for i in range(len(df.columns)) if i not in noncat_feat_idxs]
# Normalize noncategorical features
X_noncat = X[:, noncat_feat_idxs]
std_scaler = StandardScaler().fit(X_noncat)
X_noncat = std_scaler.transform(X_noncat)
X[:, noncat_feat_idxs] = X_noncat
# Run k-prototypes algorithm on all clients and obtain cluster assignment (range [1, K]) for each client
if k is None:
k = cfg['K-PROTOTYPES']['K']
k_prototypes = KPrototypes(n_clusters=k, verbose=1, n_init=cfg['K-PROTOTYPES']['N_RUNS'],
n_jobs=cfg['K-PROTOTYPES']['N_JOBS'], init='Cao', num_dissim=euclidean_dissim,
cat_dissim=matching_dissim)
client_clusters = k_prototypes.fit_predict(X, categorical=cat_feat_idxs)
k_prototypes.samples = X
k_prototypes.labels = client_clusters
k_prototypes.dist = lambda x0, x1: \
k_prototypes.num_dissim(np.expand_dims(x0[noncat_feat_idxs], axis=0), np.expand_dims(x1[noncat_feat_idxs], axis=0)) + \
k_prototypes.gamma * k_prototypes.cat_dissim(np.expand_dims(x0[cat_feat_idxs], axis=0), np.expand_dims(x1[cat_feat_idxs], axis=0))
client_clusters += 1 # Enforce that cluster labels are integer range of [1, K]
if cfg['TRAIN']['DATASET_TYPE'] == 'static_and_dynamic':
clusters_df = pd.DataFrame({'ClientID': client_ids, 'Date': dates, 'Cluster Membership': client_clusters})
clusters_df.set_index(['ClientID', 'Date'])
else:
clusters_df = pd.DataFrame({'ClientID': client_ids, 'Cluster Membership': client_clusters})
clusters_df.set_index('ClientID')
# Get centroids of clusters
cluster_centroids = np.zeros((k_prototypes.cluster_centroids_[0].shape[0],
k_prototypes.cluster_centroids_[0].shape[1] + k_prototypes.cluster_centroids_[1].shape[1]))
cluster_centroids[:, noncat_feat_idxs] = k_prototypes.cluster_centroids_[0] # Numerical features
cluster_centroids[:, cat_feat_idxs] = k_prototypes.cluster_centroids_[1] # Categorical features
#cluster_centroids = np.concatenate((k_prototypes.cluster_centroids_[0], k_prototypes.cluster_centroids_[1]), axis=1)
# Scale noncategorical features of the centroids back to original range
centroid_noncat_feats = cluster_centroids[:, noncat_feat_idxs]
centroid_noncat_feats = std_scaler.inverse_transform(centroid_noncat_feats)
cluster_centroids[:, noncat_feat_idxs] = centroid_noncat_feats
# Create a DataFrame of cluster centroids
cluster_centroids = np.rint(cluster_centroids) # Round centroids to nearest int
centroids_df = pd.DataFrame(cluster_centroids, columns=list(df.columns))
for i in range(len(data_info['SV_CAT_FEATURE_IDXS'])):
idx = data_info['SV_CAT_FEATURE_IDXS'][i]
ordinal_encoded_vals = cluster_centroids[:, idx].astype(int)
original_vals = [data_info['SV_CAT_VALUES'][idx][v] for v in ordinal_encoded_vals]
centroids_df[data_info['SV_CAT_FEATURES'][i]] = original_vals
cluster_num_series = pd.Series(np.arange(1, cluster_centroids.shape[0] + 1))
centroids_df.insert(0, 'Cluster', cluster_num_series)
# Get fraction of clients in each cluster
cluster_freqs = np.bincount(client_clusters) / float(client_clusters.shape[0])
centroids_df.insert(1, '% of Clients', cluster_freqs[1:] * 100)
# Load objects necessary for prediction and explanations
try:
scaler_ct = load(cfg['PATHS']['SCALER_COL_TRANSFORMER'])
ohe_ct_sv = load(cfg['PATHS']['OHE_COL_TRANSFORMER_SV'])
explainer = dill.load(open(cfg['PATHS']['LIME_EXPLAINER'], 'rb'))
model = load_model(cfg['PATHS']['MODEL_TO_LOAD'], compile=False)
except FileNotFoundError as not_found_err:
print('File "' + not_found_err.filename +
'" was not found. Ensure you have trained a model and run LIME before running this script.')
return
# Add model's prediction of centroids (classes and prediction probabilities) to the DataFrame
predicted_classes = []
prediction_probs = []
print("Obtaining model's predictions for cluster centroids.")
for i in tqdm(range(len(cluster_centroids))):
x = np.expand_dims(cluster_centroids[i], axis=0)
y = np.squeeze(predict_instance(x, model, ohe_ct_sv, scaler_ct).T, axis=1) # Predict centroid
prediction = 1 if y[1] >= cfg['PREDICTION']['THRESHOLD'] else 0 # Model's classification
predicted_class = cfg['PREDICTION']['CLASS_NAMES'][prediction]
predicted_classes.append(predicted_class)
prediction_probs.append(y[1] * 100) # Include as a percentage
centroids_df.insert(centroids_df.shape[1], 'At risk of chronic homelessness', pd.Series(predicted_classes))
centroids_df.insert(centroids_df.shape[1], 'Probability of chronic homelessness [%]', pd.Series(prediction_probs))
# Predict and explain the cluster centroids
if explain_centroids:
model_def = cfg['TRAIN']['MODEL_DEF'].upper()
NUM_SAMPLES = cfg['LIME'][model_def]['NUM_SAMPLES']
NUM_FEATURES = cfg['LIME'][model_def]['NUM_FEATURES']
exp_rows = []
explanations = []
print('Creating explanations for cluster centroids.')
for i in tqdm(range(cluster_centroids.shape[0])):
row = []
exp = predict_and_explain(cluster_centroids[i], model, explainer, ohe_ct_sv, scaler_ct, NUM_FEATURES, NUM_SAMPLES)
explanations.append(exp)
exp_tuples = exp.as_list()
for exp_tuple in exp_tuples:
row.extend(list(exp_tuple))
if len(exp_tuples) < NUM_FEATURES:
row.extend([''] * (2 * (NUM_FEATURES - len(exp_tuples)))) # Fill with empty space if explanation too small
exp_rows.append(row)
exp_col_names = []
for i in range(NUM_FEATURES):
exp_col_names.extend(['Explanation ' + str(i + 1), 'Weight ' + str(i + 1)])
exp_df = pd.DataFrame(exp_rows, columns=exp_col_names)
centroids_df = pd.concat([centroids_df, exp_df], axis=1, sort=False) # Concatenate client features and explanations
# Visualize clusters' LIME explanations
predictions = centroids_df[['At risk of chronic homelessness', 'Probability of chronic homelessness [%]']].to_numpy()
visualize_cluster_explanations(explanations, predictions, cluster_freqs, 'Explanations for k-prototypes clusters',
cfg['PATHS']['IMAGES'] + 'centroid_explanations_')
# Save centroid features and explanations to spreadsheet
if save_centroids:
centroids_df.to_csv(cfg['PATHS']['K-PROTOTYPES_CENTROIDS'] + datetime.now().strftime("%Y%m%d-%H%M%S") + '.csv',
index_label=False, index=False)
if save_clusters:
clusters_df.to_csv(cfg['PATHS']['K-PROTOTYPES_CLUSTERS'] + datetime.now().strftime("%Y%m%d-%H%M%S") + '.csv',
index_label=False, index=False)
return k_prototypes
def silhouette_analysis(k_min=2, k_max=20):
'''
Perform Silhouette Analysis to determine the optimal value for k. For each value of k, run k-prototypes and
calculate the average Silhouette Score. The optimal k is the one that maximizes the average Silhouette Score over
the range of k provided.
:param k_min: Smallest k to consider
:param k_max: Largest k to consider
:return: Optimal value of k
'''
cfg = yaml.full_load(open(os.getcwd() + "/config.yml", 'r'))
k_range = np.arange(k_min, k_max + 1, 1) # Range is [k_min, k_max]
silhouette_scores = []
# Run k-prototypes for each value of k in the specified range and calculate average Silhouette Score
for k in k_range:
print('Running k-prototypes with k = ' + str(k))
clustering = cluster_clients(k=k, save_centroids=False, save_clusters=False, explain_centroids=False)
print('Calculating average Silhouette Score for k = ' + str(k))
silhouette_avg = silhouette_score(clustering.samples, clustering.labels, metric=clustering.dist, sample_size=1500)
silhouette_scores.append(silhouette_avg)
optimal_k = k_range[np.argmax(silhouette_scores)] # Optimal k is that which maximizes average Silhouette Score
# Visualize the Silhouette Plot
visualize_silhouette_plot(k_range, silhouette_scores, optimal_k=optimal_k,
file_path=cfg['PATHS']['IMAGES'] + 'silhouette_plot_')
return optimal_k
if __name__ == '__main__':
cfg = yaml.full_load(open("./config.yml", 'r'))
if cfg['K-PROTOTYPES']['EXPERIMENT'] == 'cluster_clients':
_ = cluster_clients(k=None, save_centroids=True, save_clusters=True, explain_centroids=True)
else:
optimal_k = silhouette_analysis(k_min=cfg['K-PROTOTYPES']['K_MIN'], k_max=cfg['K-PROTOTYPES']['K_MAX'])