Spaces:
Runtime error
Runtime error
File size: 20,467 Bytes
d2a8669 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 |
## Todo utils for data pre-process
import pandas as pd
import numpy as np
import networkx as nx
import scipy.sparse as sp
import re
from alibaba_processing.ali_CatGCN_pre_processing import get_count, filter_triplets, col_map, label_map
def load_networkx_file(model_type, data_extension, dataset_name, dataset_path, dataset_user_id_name, onehot_bin_columns, onehot_cat_columns, sens_attr, predict_attr):
# load data from graphml to csv
#print('Loading dataset for FairGNN...')
#print(data_extension)
print('Extracting networkx data format...')
if data_extension == '.graphml':
data = nx.read_graphml(dataset_path)
elif data_extension == '.gexf':
data = nx.read_gexf(dataset_path)
elif data_extension == '.gml':
data = nx.read_gml(dataset_path)
elif data_extension == '.leda':
data = nx.read_leda(dataset_path)
elif data_extension == '.net':
data = nx.read_pajek(dataset_path)
# load graph nodes
#print('Data extension', data_extension)
#print('Data', data)
df_nodes = pd.DataFrame.from_dict(dict(data.nodes(data=True)), orient='index')
# check if user_id column is not assigned as the index
if df_nodes.columns[0] != dataset_user_id_name:
# if so, then we make it as the first column
df_nodes = df_nodes.reset_index(level=0)
df_nodes = df_nodes.rename(columns={"index": dataset_user_id_name})
# check if user_id column is not string
if type(df_nodes[dataset_user_id_name][0]) != np.int64:
# if so, we convert it to int
df_nodes[dataset_user_id_name] = pd.to_numeric(df_nodes[dataset_user_id_name])
df_nodes = df_nodes.astype({dataset_user_id_name: int})
#check if sens_attr and predict_attr is float
if (df_nodes[sens_attr].dtype == np.float64):
df_nodes[sens_attr] = df_nodes[sens_attr].astype(int)
if (df_nodes[predict_attr].dtype == np.float64):
df_nodes[predict_attr] = df_nodes[predict_attr].astype(int)
# todo if dataset will be used for RHGN or CatGCN then return, else we assume for FairGNN then complete the onehot encoding process
if model_type == 'RHGN':
return df_nodes
elif model_type == 'CatGCN':
if dataset_name == 'nba' or dataset_name == 'pokec_z' or dataset_name == 'pokec_n':
df_edge_list = nx.to_pandas_edgelist(data)
return df_nodes, df_edge_list
else: ## Data is either Alibaba or tecent and they will get their edges later on
df_edge_list = None
return df_nodes, df_edge_list
else: # FairGNN
if dataset_name == 'alibaba' or dataset_name == 'tecent':
if dataset_name == 'tecent':
df_nodes = bin_age_range_tecent(df_nodes)
df_nodes = df_nodes.drop(columns=["cid1_name", "cid2_name ", "cid3_name", "item_name", "seg_name"])
if dataset_name == 'alibaba':
df_nodes = bin_alibaba(df_nodes)
edges_path = create_edges(df_nodes, dataset_name)
df_edge_list = edges_path
# todo add one-hot encoding
# add binary onehot encoding if needed
if onehot_bin_columns is not None:
df_nodes = apply_bin_columns(df_nodes, onehot_bin_columns)
# add categorical onehot encoding if needed
if onehot_cat_columns is not None:
df_nodes = apply_cat_columns(df_nodes, onehot_cat_columns)
if dataset_name == 'nba' or dataset_name == 'pokec':
# load graph edges
df_edge_list = nx.to_pandas_edgelist(data)
#save the edges as .txt file
edges_path = './FairGNN_data_relationship'
df_edge_list.to_csv(r'{}.txt'.format(edges_path), header=None, index=None, sep=' ', mode='a')
return df_nodes, edges_path
def load_neo4j_file(model_type, dataset_path, dataset_name, uneeded_columns, onehot_bin_columns, onehot_cat_columns):
# todo pre-process node and edge data
#print('Loading dataset for FairGNN...')
print('Extracting neo4j data format...')
df = pd.read_json(dataset_path, lines=True) # may cause error
# todo extract node csv
nodes_df = df.loc(df['type'] == ['node'])
#delete un-needed column
nodes_df = nodes_df.drop(['label', 'start', 'end'], axis=1)
# get nodes properties as list of json
prop_list = []
id_list = []
labels_list = []
for index, row in nodes_df.iterrows():
prop_list.append(row['propertiees'])
id_list.append(row['id'])
labels_list.append(row['labels'])
for i in range(len(prop_list)):
prop_list[i]['id'] = id_list[i]
prop_list[i]['labels'] = labels_list[i]
# create new csv from the prop list
new_nodes_df = pd.DataFrame(prop_list)
new_nodes_df = new_nodes_df.drop(['properties'], axis=1)
# make id as first column
first_column = new_nodes_df.pop('id')
new_nodes_df.insert(0, 'id', first_column)
# we only apply the uneeded columns feature and the onehot encoding for the the FairGNN model
if model_type == 'FairGNN':
# add binary onehot encoding if needed
if onehot_bin_columns is not None:
new_nodes_df = apply_bin_columns(new_nodes_df, onehot_bin_columns)
# add categorical onehot encoding if needed
if onehot_cat_columns is not None:
new_nodes_df = apply_cat_columns(new_nodes_df, onehot_cat_columns)
# todo remove columns that we don't want to have in the dataframe
if len(uneeded_columns) == 0:
new_nodes_df = remove_column_from_df('description') ## we don't want descriptions in our code per default
else:
new_nodes_df = remove_column_from_df(uneeded_columns) ## user defined columns
# now we remove columns that we don't want it to change for the next step (one-hot step) (e.g. id, person id)
new_nodes_df = remove_unneeded_columns(new_nodes_df)
# replace nan with 0
new_nodes_df = new_nodes_df.replace(r'^\s*$', np.nan, regex=True)
new_nodes_df = new_nodes_df.fillna(0)
# Todo know which columns to filter out
# not needed -- replacment the function apply_cat_columns
#new_nodes_df = apply_one_hot_encodding(nodes_columns, new_nodes_df)
############################################
#extract edges relationships
if dataset_name == 'alibaba' or dataset_name == 'tecent':
return new_nodes_df
else:
edges_df = df.loc[(df['type'] == 'relationship')]
edges_df = edges_df.drop(['labels'], axis=1)
edges_relation = pd.DataFrame(columns=['start', 'end'], index=range(len(edges_df.index)))
i = 0
for index, row in edges_df.iterrows():
edges_relation['start'][i] = row['start']['id']
edges_relation['end'][i] = row['end']['id']
i = i+1
edges_relation.columns = [''] * len(edges_relation.columns)
# save .txt
# todo maybe return it normally?
edges_path = './FairGNN_data_relationship'
edges_relation.to_csv(r'{}.txt'.format(edges_path), sep='\t', header=False, index=False)
return new_nodes_df, edges_relation
def remove_column_from_df(column, df):
nodes_columns = df.columns.tolist()
# check if we have list of columns or not
if type(column) == list:
for i in column:
df = df.drop([i], axis=1)
else:
for c in nodes_columns:
if c == column:
df = df.drop([column], axis=1)
def remove_unneeded_columns(new_nodes_df):
unneeded_columns = []
nodes_columns = new_nodes_df.columns.tolist()
matchers = ['id', 'iD', 'Id', 'name']
matching = [s for s in nodes_columns if any(xs in s for xs in matchers)]
for i in range(len(matching)):
if matching[i].endswith('id') or matching[i].endswith('Id'):
unneeded_columns.append(matching[i])
nodes_columns.remove(matching|[i])
if matching[i] == 'name':
nodes_columns.remvoe(matching[i])
nodes_columns.remove('id')
nodes_columns.remove('labels')
return nodes_columns
def apply_one_hot_encodding(nodes_columns, new_nodes_df):
for column in nodes_columns:
if new_nodes_df[column].dtype != 'int64' or new_nodes_df[column].dtype != 'float64':
new_nodes_df[column] = new_nodes_df[column].apply(lambda x: ",".join(x) if isinstance(x, list) else x)
tempdf = pd.get_dummies(new_nodes_df[column], prefix=column, drop_first=True)
new_nodes_df = pd.merge(left=new_nodes_df, right=tempdf, left_index=True, right_index=True)
new_nodes_df = new_nodes_df.drop(columns=column)
new_nodes_df.columns = new_nodes_df.columns.str.replace(' \t', '')
new_nodes_df.columns = new_nodes_df.columns.str.strip().str.replace(' ', '_')
new_nodes_df.columns = new_nodes_df.columns.str.replace('___', '_')
new_nodes_df.columns = new_nodes_df.columns.str.replace('__', '_')
return new_nodes_df
def fair_metric(output,idx, labels, sens):
#output == target
val_y = labels[idx].cpu().numpy()
idx_s0 = sens.cpu().numpy()[idx.cpu().numpy()]==0
idx_s1 = sens.cpu().numpy()[idx.cpu().numpy()]==1
# parameters for "overall accuracy equality"
#true_y = np.asarray(output)
#true_y = output.detach().numpy()
#true_y = np.asarray(true_y)
# Use tensor.detach().numpy()
#y0_s0 = np.bitwise_and(true_y == 0, idx_s0)
#y0_s1 = np.bitwise_and(true_y == 0, idx_s1)
#y1_s0 = np.bitwise_and(true_y == 1, idx_s0)
#y1_s1 = np.bitwise_and(true_y == 1, idx_s1)
idx_s0_y1 = np.bitwise_and(idx_s0,val_y==1)
idx_s1_y1 = np.bitwise_and(idx_s1,val_y==1)
idx_s0_y0 = np.bitwise_and(idx_s0,val_y==0)
idx_s1_y0 = np.bitwise_and(idx_s1,val_y==0)
pred_y = (output[idx].squeeze()>0).type_as(labels).cpu().numpy()
#parity = abs(sum(pred_y[idx_s0])/sum(idx_s0)-sum(pred_y[idx_s1])/sum(idx_s1))
parity = np.abs(sum(pred_y[idx_s0])/sum(idx_s0)-sum(pred_y[idx_s1])/sum(idx_s1))
print('parity debug')
print('pred_y:',pred_y)
print('pred_y[idx_s0]:', pred_y[idx_s0])
print('idx_s0:', idx_s0)
print('parity:', parity)
#equality = abs(sum(pred_y[idx_s0_y1])/sum(idx_s0_y1)-sum(pred_y[idx_s1_y1])/sum(idx_s1_y1))
equality = np.abs(sum(pred_y[idx_s0_y1])/sum(idx_s0_y1)-sum(pred_y[idx_s1_y1])/sum(idx_s1_y1))
# treatment equality
te1_s0 = (sum(pred_y[idx_s0_y0]) / sum(idx_s0_y0)) / (np.count_nonzero(pred_y[idx_s0_y1] == 0) / sum(idx_s0_y1))
te1_s1 = (sum(pred_y[idx_s1_y0]) / sum(idx_s1_y0)) / (np.count_nonzero(pred_y[idx_s1_y1] == 0) / sum(idx_s1_y1))
te_diff_1 = te1_s0 - te1_s1
abs_ted_1 = abs(te_diff_1)
te0_s0 = (np.count_nonzero(pred_y[idx_s0_y1] == 0) / sum(idx_s0_y1)) / (sum(pred_y[idx_s0_y0]) / sum(idx_s0_y0))
te0_s1 = (np.count_nonzero(pred_y[idx_s1_y1] == 0) / sum(idx_s1_y1)) / (sum(pred_y[idx_s1_y0]) / sum(idx_s1_y0))
te_diff_0 = te0_s0 - te0_s1
abs_ted_0 = abs(te_diff_0)
if abs_ted_0 < abs_ted_1:
te_s0 = te0_s0
te_s1 = te0_s1
te_diff = te_diff_0
else:
te_s0 = te1_s0
te_s1 = te1_s1
te_diff = te_diff_1
# "overall accuracy equality"
oae_s0 = np.count_nonzero(pred_y[idx_s0_y0] == 0) / sum(idx_s0_y0) + sum(pred_y[idx_s0_y1]) / sum(idx_s0_y1)
oae_s1 = np.count_nonzero(pred_y[idx_s1_y0] == 0) / sum(idx_s1_y0) + sum(pred_y[idx_s1_y1]) / sum(idx_s1_y1)
oae_diff = np.abs(oae_s0 - oae_s1)
# disparate_impact
return parity, equality,oae_diff, te_diff
def apply_bin_columns(df, onehot_bin_columns):
for column in df:
if column in onehot_bin_columns:
df[column] = df[column].astype(int)
return df
def apply_cat_columns(df, onehot_cat_columns):
df = pd.get_dummies(df, columns=onehot_cat_columns)
return df
def create_edges(df_nodes, dataset_name):
if dataset_name == 'alibaba':
# divide data
df_user = df_nodes[['userid', 'final_gender_code', 'age_level', 'pvalue_level', 'occupation', 'new_user_class_level ']].copy()
df_item = df_nodes[['adgroup_id', 'cate_id']].copy()
df_click = df_nodes[['userid', 'adgroup_id', 'clk']].copy()
df_user.dropna(inplace=True)
df_user.rename(columns={'userid':'uid', 'final_gender_code':'gender','age_level':'age', 'pvalue_level':'buy', 'occupation':'student', 'new_user_class_level ':'city'}, inplace=True)
df_item.rename(columns={'adgroup_id':'pid','cate_id':'cid'}, inplace=True)
df_click.rename(columns={'userid':'uid','adgroup_id':'pid'}, inplace=True)
df_click = df_click[df_click['clk']>0]
df_click.drop('clk', axis=1, inplace=True)
df_click = df_click[df_click['uid'].isin(df_user['uid'])]
df_click = df_click[df_click['pid'].isin(df_click['pid'])]
df_click.drop_duplicates(inplace=True)
uid_pid, uid_activity, pid_popularity = filter_triplets(df_click, 'uid', 'pid', min_uc=0, min_sc=2) # min_sc>=2
#sparsity = 1. * uid_pid.shape[0] / (uid_activity.shape[0] * pid_popularity.shape[0])
uid_pid_cid = pd.merge(uid_pid, df_item, how='inner', on='pid')
raw_uid_cid = uid_pid_cid.drop('pid', axis=1, inplace=False)
raw_uid_cid.drop_duplicates(inplace=True)
uid_cid, uid_activity, cid_popularity = filter_triplets(raw_uid_cid, 'uid', 'cid', min_uc=0, min_sc=2) # min_sc>=2
#sparsity = 1. * uid_cid.shape[0] / (uid_activity.shape[0] * cid_popularity.shape[0])
uid_pid = uid_pid[uid_pid['uid'].isin(uid_cid['uid'])]
uid_pid_1 = uid_pid[['uid','pid']].copy()
uid_pid_1.rename(columns={'uid':'uid1'}, inplace=True)
uid_pid_2 = uid_pid[['uid','pid']].copy()
uid_pid_2.rename(columns={'uid':'uid2'}, inplace=True)
uid_pid_uid = pd.merge(uid_pid_1, uid_pid_2, how='inner', on='pid')
uid_uid = uid_pid_uid.drop('pid', axis=1, inplace=False)
uid_uid.drop_duplicates(inplace=True)
del uid_pid_1, uid_pid_2, uid_pid_uid
# map
user_label = df_user[df_user['uid'].isin(uid_cid['uid'])]
uid2id = {num: i for i, num in enumerate(user_label['uid'])}
cid2id = {num: i for i, num in enumerate(pd.unique(uid_cid['cid']))}
user_label = col_map(user_label, 'uid', uid2id)
user_label = label_map(user_label, user_label.columns[1:])
user_edge = uid_uid[uid_uid['uid1'].isin(uid_cid['uid'])]
user_edge = user_edge[user_edge['uid2'].isin(uid_cid['uid'])]
user_edge = col_map(user_edge, 'uid1', uid2id)
user_edge = col_map(user_edge, 'uid2', uid2id)
return user_edge
elif dataset_name == 'tecent':
df_user = df_nodes[['user_id', 'gender', 'age_range']].copy()
df_user.dropna(inplace=True)
df_user.rename(columns={"user_id":"uid", "age_range":"age"}, inplace=True)
df_item = df_nodes[['item_id', 'cid3']].copy()
df_item.dropna(inplace=True)
df_item.rename(columns={"item_id":"pid", "cid3":"cid"}, inplace=True)
df_item.reset_index(drop=True, inplace=True)
df_click = df_nodes[['user_id', 'item_id']].copy()
df_click.dropna(inplace=True)
df_click.rename(columns={"user_id":"uid", "item_id":"pid"}, inplace=True)
df_click.reset_index(drop=True, inplace=True)
df_item = df_item.sample(frac=0.15, random_state=11)
df_item.reset_index(drop=True, inplace=True)
df_click = df_click.sample(frac=0.15, random_state=11)
df_click.reset_index(drop=True, inplace=True)
df_click = df_click[df_click["uid"].isin(df_user["uid"])]
df_click = df_click[df_click["pid"].isin(df_item["pid"])]
df_click.drop_duplicates(inplace=True)
df_click.reset_index(drop=True, inplace=True)
df_click, uid_activity, pid_popularity = filter_triplets(df_click, 'uid', 'pid', min_uc=0, min_sc=2)
sparsity = 1. * df_click.shape[0] / (uid_activity.shape[0] * pid_popularity.shape[0])
df_click_item = pd.merge(df_click, df_item, how="inner", on="pid")
raw_click_item = df_click_item.drop("pid", axis=1, inplace=False)
raw_click_item.drop_duplicates(inplace=True)
df_click_item, uid_activity, cid_popularity = filter_triplets(raw_click_item, 'uid', 'cid', min_uc=0, min_sc=2)
sparsity = 1. * df_click_item.shape[0] / (uid_activity.shape[0] * cid_popularity.shape[0])
df_click = df_click[df_click["uid"].isin(df_click_item["uid"])]
df_click_1 = df_click[["uid", "pid"]].copy()
df_click_1.rename(columns={"uid":"uid1"}, inplace=True)
df_click_2 = df_click[["uid", "pid"]].copy()
df_click_2.rename(columns={"uid":"uid2"}, inplace=True)
df_click1_click2 = pd.merge(df_click_1, df_click_2, how="inner", on="pid")
df_uid_uid = df_click1_click2.drop("pid", axis=1, inplace=False)
df_uid_uid.drop_duplicates(inplace=True)
del df_click_1, df_click_2, df_click1_click2
# map
df_label = df_user[df_user["uid"].isin(df_click_item["uid"])]
uid2id = {num: i for i, num in enumerate(df_label['uid'])}
cid2id = {num: i for i, num in enumerate(pd.unique(df_click_item['cid']))}
df_label = col_map(df_label, 'uid', uid2id)
df_label = label_map(df_label, df_label.columns[1:])
user_edge = df_uid_uid[df_uid_uid['uid1'].isin(df_click_item['uid'])]
user_edge = user_edge[user_edge['uid2'].isin(df_click_item['uid'])]
user_edge = col_map(user_edge, 'uid1', uid2id)
user_edge = col_map(user_edge, 'uid2', uid2id)
return user_edge
def bin_age_range_tecent(df_nodes):
age_dic = {'11~15':0, '16~20':0, '21~25':0, '26~30':1, '31~35':1, '36~40':2, '41~45':2, '46~50':3, '51~55':3, '56~60':4, '61~65':4, '66~70':4, '71~':4}
df_nodes[["age_range"]] = df_nodes[["age_range"]].applymap(lambda x:age_dic[x])
#df_nodes.rename(columns={"user_id":"uid", "age_range":"age"}, inplace=True)
#df_nodes["bin_age"] = df_nodes["age"]
df_nodes["age_range"] = df_nodes["age_range"].replace(1,0)
df_nodes["age_range"] = df_nodes["age_range"].replace(2,1)
df_nodes["age_range"] = df_nodes["age_range"].replace(3,1)
df_nodes["age_range"] = df_nodes["age_range"].replace(4,1)
return df_nodes
def bin_alibaba(df_nodes):
df_nodes["age_level"] = df_nodes["age_level"].replace(1,0)
df_nodes["age_level"] = df_nodes["age_level"].replace(2,0)
df_nodes["age_level"] = df_nodes["age_level"].replace(3,0)
df_nodes["age_level"] = df_nodes["age_level"].replace(4,1)
df_nodes["age_level"] = df_nodes["age_level"].replace(5,1)
df_nodes["age_level"] = df_nodes["age_level"].replace(6,1)
df_nodes['pvalue_level'] = df_nodes['pvalue_level'].replace(3.0, 2.0)
df_nodes['pvalue_level'] = df_nodes['pvalue_level'].astype('int64')
return df_nodes
def calculate_dataset_fairness(df, dataset_name, sens_attr, label):
if dataset_name == 'pokec_z':
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(-1, 0)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(0, 0)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(1, 0)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(2, 1)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(3, 1)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(4, 1)
elif dataset_name == 'pokec_n':
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(-1, 0)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(0, 1)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(1, 1)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(2, 1)
df['I_am_working_in_field'] = df['I_am_working_in_field'].replace(3, 1)
total_number_of_sens0 = len(df.loc[df[sens_attr] == 0])
total_number_of_sens1 = len(df.loc[df[sens_attr] == 1])
number_of_positive_sens0 = len(df.loc[(df[sens_attr] == 0) & (df[label] == 1)])
number_of_positive_sens1 = len(df.loc[(df[sens_attr] == 1) & (df[label] == 1)])
fairness = np.absolute(number_of_positive_sens0) / np.absolute(total_number_of_sens0) - np.absolute(number_of_positive_sens1) / np.absolute(total_number_of_sens1)
return fairness * 100
|