Blame view
MLTD/src/PredictionKorvesis.py
20.7 KB
0d8c0f816
initial commit
|
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 |
# -*- coding: utf-8 -*-
import multiprocessing
import pandas as pd
import numpy as np
import math
from sklearn.ensemble import RandomForestRegressor
import sklearn_relief as rrelief
import traceback
import sys
import logging
#suppress warning from insert row method line:df1.loc[row_number] = row_value
import warnings
import Utils
warnings.filterwarnings("ignore", category=UserWarning)
#suppress warning from pandas
pd.options.mode.chained_assignment = None # default='warn'
logger = logging.getLogger(__name__)
def sigmoid(s, midpoint, t, ep_length):
return (1 / (1 + math.exp(s * (ep_length - midpoint - t))))
def preprocess_data_binarize(
dataset, chunks_type, unique_event_ids=[], check_unknown_events=False
):
if len(unique_event_ids) == 0:
unique_event_ids = dataset["Event_id"].unique()
unique_event_ids = unique_event_ids[unique_event_ids != -1]
# group by hour and merge all the event ids to a list
dataset = dataset.groupby(
pd.to_datetime(dataset["Timestamps"]).dt.floor(chunks_type)
)["Event_id"].apply(set)
# create a 2d array (the final dataset) with zeros and name the columns to
# event types
df = pd.DataFrame(
np.zeros(shape=(len(dataset), len(unique_event_ids))), columns=unique_event_ids
)
df["Timestamps"] = dataset.index.values
if check_unknown_events:
for i in range(len(dataset)):
if len(dataset[i]) == 0:
continue
for event in dataset[i]:
if event in df.columns:
df.loc[i, event] = 1
else:
for i in range(len(dataset)):
df.loc[i, dataset[i]] = 1
return df
def preprocess_data_remove_rare_events(df, target_event, remove_threshold):
"""
Remove the event types with frequency less than
remove_threshold*freq(target_event) or remove_threshold*mean_freq.
Parameters
----------
df : dataframe
dataset with the events
target_event : int
event id of the target event
remove_threshold : float
percentage of the frequency of the target event below of which the
events are removed
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if not df_valid(df):
logger.error(f"RRE df not valid")
raise
events_freq = df[
df.columns[~df.columns.isin(["Timestamps", "Timedeltas", target_event])]
].sum()
te_freq = df.loc[:, df.columns == target_event].sum()[target_event]
mean_freq = events_freq.mean()
rare_freq = te_freq * remove_threshold
if te_freq > mean_freq:
logger.warning(
f"RRE: Target event frequent "
f"{te_freq} "
f"is higher than the mean frequency of the rest of the events "
f"{mean_freq}"
)
if rare_freq > mean_freq:
logger.warning(
f"RRE: Rare event frequent "
f"{rare_freq} "
f"based on target event frequency "
f"{te_freq}*{remove_threshold} "
f"is higher than the mean frequency "
f"{mean_freq}"
f", changing rare threshold to mean based "
f"{mean_freq}*{remove_threshold}={mean_freq*remove_threshold}"
)
rare_freq = mean_freq * remove_threshold
remove_candidate_events = events_freq[events_freq < rare_freq].index
logger.info(
f"RRE: Removing "
f"{len(remove_candidate_events)} "
f"events out of the "
f"{len(events_freq)} "
f"total events "
)
df = df.drop(columns=remove_candidate_events)
df = remove_empty_episodes(df, target_event)
if not df_valid(df):
logger.error(f"RRE parameters too strict")
raise
return df
def remove_empty_episodes(df, target_event):
"""
Remove the episodes with no rows.
Parameters
----------
df : np.DataFrame
The dataset
target_event : int
The target event id
Returns
-------
df : np.DataFrame
The cleaned dataset.
"""
remove_candidates = []
te_indeces = df.index[df[target_event] == 1].tolist()
for i in range(len(te_indeces) - 1):
if te_indeces[i] + 1 == te_indeces[i + 1]:
remove_candidates.append(te_indeces[i + 1])
if len(remove_candidates) > 0:
logger.info(
f"CLEAN: Removed {len(remove_candidates)} empty episodes"
f" remaining {len(te_indeces)-len(remove_candidates)} episodes"
)
df.drop(remove_candidates, inplace=True)
df.set_index(pd.Series(range(len(df))), inplace=True)
return df
def preprocess_data_remove_frequent_events(df, target_event, remove_threshold):
"""
Remove the event types with frequency more than remove_threshold*max_freq.
Parameters
----------
df : dataframe
dataset with the events
target_event : int
event id of the target event
remove_threshold : float
percentage of the frequency of the target event below of which the
events are removed
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if not df_valid(df):
logger.error(f"RFE df not valid")
raise
events_freq = df[
df.columns[~df.columns.isin(["Timestamps", "Timedeltas", target_event])]
].sum()
max_freq_event = events_freq.idxmax()
if str(max_freq_event) == str(target_event):
logger.warning(
f"RFE: Target event is the most frequent "
f"{events_freq.loc[target_event]} "
f"no meaning in removing frequent events. "
f"Remove frequency won't run"
)
return df
else:
max_freq = events_freq.loc[max_freq_event]
remove_candidate_events = events_freq[
events_freq >= max_freq - (max_freq * remove_threshold)
].index
if len(remove_candidate_events) >= len(df.columns) - 4:
logger.info(
f"RFE: No meaning in applying RFE as all the events will be removed. "
f"{len(remove_candidate_events)} "
f"events out of the "
f"{len(df.columns) - 4} "
f"total events "
)
return df
logger.info(
f"RFE: Maximum frequency is "
f"{max_freq}. "
f"Removing "
f"{len(remove_candidate_events)}"
f" events with frequency >= "
f"{max_freq - (max_freq * remove_threshold)}"
)
if target_event in remove_candidate_events:
remove_candidate_events = remove_candidate_events.drop(target_event)
df = df.drop(columns=remove_candidate_events)
df = remove_empty_episodes(df, target_event)
if not df_valid(df):
logger.error(f"RFE parameters too strict")
raise
return df
def preprocess_data_keep_only_first_event(df, target_event):
"""
Keeps only the first occurence of events per episode
(i.e. episodes = between target_event occurences).
Parameters
----------
df : dataframe
dataset with the events
target_event : int
event id of the target event
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if not df_valid(df):
logger.error(f"KOFE df not valid")
raise
start_row_index = 0
removed_events_cnt = 0
for index, row in df.iterrows():
if row[target_event] == 1:
# slice the df to the episode of the target event
# (excluding the row of the target even)
# change index to index+1 if you want to include the target event
# row
episode = df.iloc[
start_row_index:index,
]
episode = episode[
episode.columns[
~episode.columns.isin(["Timestamps", "Timedeltas", target_event])
]
]
for column_index in episode:
column = episode[column_index]
for i in range(
len(column) - 1, 0, -1
): # iterating bottom up to search for trailing 1s
if column.iloc[i] == 1 and column.iloc[i] == column.iloc[i - 1]:
actual_row_index = start_row_index + i
df.loc[actual_row_index, column_index] = 0
removed_events_cnt += 1
if index + 1 < len(df): # start from the row after the target event
start_row_index = index + 1
if removed_events_cnt > 0:
logger.info(f"KOFE: Removed {removed_events_cnt} events")
df = remove_empty_episodes(df, target_event)
if not df_valid(df):
logger.error(f"KOFE parameters too strict")
raise
return df
def compute_risk(df, target_event, s, midpoint, buffer_time=None, risk_ignore_buffer=True):
"""
Compute the risk based on the Sigmoid function.
Parameters
----------
df : pd.DataFrame
Dataset with the events
target_event : int
event id of the target event
ep_length : str
The episode length expressed with a string e.g. '8H'
s : float
steepness
midpoint : str
skewness expressed with a string e.g. '8H'
buffer_time : str
The buffered time before the target event to which the risk should be
zero, expressed with a string e.g. '1H'
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if "Timedeltas" not in df.columns:
df["Timedeltas"] = compute_timedeltas_mins(df, target_event)
midpoint_mins = Utils.convert_hours_to_mins(Utils.strtime_to_hours(midpoint))
buffer_time_mins = 0
if buffer_time is not None:
buffer_time_mins = Utils.convert_hours_to_mins(Utils.strtime_to_hours(buffer_time))
start_row_index = 0
for index, row in df.iterrows():
if row[target_event] == 1:
# slice the df to the episode of the target event (including the row of the target event)
# change index+1 to index if you want to exclude the target event row
episode = df.iloc[
start_row_index : (index + 1),
]
for seg_index, ep_row in episode.iterrows():
segment_time_mins = episode.loc[seg_index, "Timedeltas"]
if buffer_time is not None and segment_time_mins <= buffer_time_mins:
df.loc[seg_index, "Risk"] = 0
else:
if buffer_time is not None and risk_ignore_buffer:
segment_time_mins = segment_time_mins - buffer_time_mins
df.loc[seg_index, "Risk"] = Utils.sigmoid_risk(
s, midpoint_mins, segment_time_mins
)
start_row_index = index + 1
return df
def insert_row(row_number, df, row_value):
"""
Function to insert row in the dataframe
Parameters
----------
row_number : int
the index of the row to insert
df : dataframe
the dataframe
row_value : int
the value to insert
Returns
-------
df : np.DataFrame
The dataset with the inserted row.
"""
# Slice the upper half of the dataframe
df1 = df[0:row_number]
# Store the result of lower half of the dataframe
df2 = df[row_number:]
# Insert the row in the upper half dataframe
df1.loc[row_number] = row_value
# Concat the two dataframes
df_result = pd.concat([df1, df2])
# Reassign the index labels
df_result.index = [*range(df_result.shape[0])]
return df_result
def binary_sum(x):
if x.sum() > 0:
return 1.0
else:
return 0.0
def df_valid(df):
return len(df) > 0 and len(df.columns) > 4
def preprocess_data_MIL_oversampling(df, target_event, window_size, threshold):
"""
Apply Multi-Instance Learning oversampling the events closer to the target
event. The merged row is added right below the window from which is computed
Parameters
----------
df : dataframe
dataset with the events
target_event : int
event id of the target event
window_size : int
the window size of the sampling
threshold : float
specifies when the oversampling begins
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if not df_valid(df):
logger.error(f"MIL_OVER df not valid")
raise
new_rows = []
start_row_index = 0
for index, row in df.iterrows():
if row[target_event] == 1:
# slice the df to the episode of the target event (excluding the
# row of the target even)
# change index to index+1 if you want to include the target event row
episode = df.iloc[
start_row_index:index,
]
ep_index = 0
for row_index, ep_row in episode.iterrows():
if (
row_index + window_size >= index + 1
): # stop right before the target event
break
if ep_row["Risk"] >= threshold:
window = episode.iloc[
ep_index : (ep_index + window_size),
]
new_row = window[
window.columns[
~window.columns.isin(["Timestamps", "Timedeltas", "Risk"])
]
].apply(binary_sum, axis=0)
new_row["Risk"] = window.tail(1)["Risk"].item()
new_row["Timestamps"] = np.datetime_as_string(
window.tail(1)["Timestamps"], unit="s"
)[
0
] # convert nanosec to string Datetime
new_row["Timedeltas"] = window.tail(1)["Timedeltas"].item()
new_rows.append(
(row_index + window_size - 1 + len(new_rows), new_row)
) # row_index is the index of the row of the df
ep_index += 1
if index + 1 < len(df): # start from the row after the target event
start_row_index = index + 1
logger.info(
f"MIL_OVER: Adding {len(new_rows)} new/merged occurrences " f"of events"
)
for row_index, values in new_rows:
df = insert_row(row_index, df, values)
return df
def preprocess_data_feature_selection(df, target_event):
"""
Apply Feature Selection and more specifically the RRelief algorithm.
It keeps only the events with non-zero weight.
Parameters
----------
df : dataframe
dataset with the events
target_event : int
event id of the target event
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if not df_valid(df):
logger.error(f"FS df not valid")
raise
r = rrelief.Relief()
input_matrix = df.iloc[:, :-3].to_numpy()
label_vector = df.iloc[:, :-3].columns.values
transformed_matrix = r.fit_transform(input_matrix, df["Risk"].array)
weights = dict(zip(label_vector, r.w_))
fs = {k: v for k, v in weights.items() if v > 0}
# fs = dict(sorted(weights.items(), key=lambda x: x[1], reverse=True)[:top])
keep_column_labels = [*fs.keys()]
if target_event not in fs.keys():
keep_column_labels.append(target_event)
keep_column_labels.append("Risk")
keep_column_labels.append("Timestamps")
keep_column_labels.append("Timedeltas")
df = df[keep_column_labels]
if len(weights) - len(fs) > 0:
logger.info(
f"FS: Removed {len(weights)-len(fs)} out of {len(weights)} features"
)
if not df_valid(df):
logger.error(f"FS parameters too strict")
raise
return df
def compute_risk(df, target_event, s, midpoint, buffer_time=None, risk_ignore_buffer=True):
"""
Compute the risk based on the Sigmoid function.
Parameters
----------
df : pd.DataFrame
Dataset with the events
target_event : int
event id of the target event
ep_length : str
The episode length expressed with a string e.g. '8H'
s : float
steepness
midpoint : str
skewness expressed with a string e.g. '8H'
buffer_time : str
The buffered time before the target event to which the risk should be
zero, expressed with a string e.g. '1H'
Returns
-------
df : np.DataFrame
The preprocessed dataset.
"""
if "Timedeltas" not in df.columns:
df["Timedeltas"] = compute_timedeltas_mins(df, target_event)
midpoint_mins = Utils.convert_hours_to_mins(Utils.strtime_to_hours(midpoint))
buffer_time_mins = 0
if buffer_time is not None:
buffer_time_mins = Utils.convert_hours_to_mins(Utils.strtime_to_hours(buffer_time))
start_row_index = 0
for index, row in df.iterrows():
if row[target_event] == 1:
# slice the df to the episode of the target event (including the row of the target event)
# change index+1 to index if you want to exclude the target event row
episode = df.iloc[
start_row_index : (index + 1),
]
for seg_index, ep_row in episode.iterrows():
segment_time_mins = episode.loc[seg_index, "Timedeltas"]
if buffer_time is not None and segment_time_mins <= buffer_time_mins or index == seg_index:
df.loc[seg_index, "Risk"] = 0
else:
if buffer_time is not None and risk_ignore_buffer:
segment_time_mins = segment_time_mins - buffer_time_mins
df.loc[seg_index, "Risk"] = Utils.sigmoid_risk(
s, midpoint_mins, segment_time_mins
)
start_row_index = index + 1
return df
def compute_timedeltas_mins(df, target_event):
start_row_index = 0
dataset_timedeltas = []
for index, row in df.iterrows():
if row[target_event] == 1:
# slice the df to the episode of the target event (excluding the
# row of the target even)
# change index to index+1 if you want to include the target event row
episode = df.iloc[
start_row_index : index + 1,
]
te_date = episode["Timestamps"].values[-1]
for timestamp in episode["Timestamps"].values:
timedelta = te_date - timestamp
time_distance_mins = timedelta.astype('timedelta64[m]') / np.timedelta64(1, 'm')
dataset_timedeltas.append(time_distance_mins)
if index + 1 < len(df): # start from the row after the target event
start_row_index = index + 1
return dataset_timedeltas
def train_rf_model(dataset, chunks_type, target_event, s, midpoint, rre=False, rfe=False, kofe=False, mil_over=False,
mil_down=False, fs=False, rre_thres=0.2,rfe_thres=0.1,mil_thres=0.8,mil_window_size=2, fs_top=20,
rf_n_jobs=multiprocessing.cpu_count(), rf_max_depth=12, rf_random_state=0,rf_n_estimators=1000):
try:
df = preprocess_data_binarize(dataset, chunks_type)
df["Timedeltas"] = compute_timedeltas_mins(df, target_event)
if rre:
df = preprocess_data_remove_rare_events(df,target_event,rre_thres)
if rfe:
df = preprocess_data_remove_frequent_events(df,target_event,rfe_thres)
if kofe:
df = preprocess_data_keep_only_first_event(df, target_event)
df = compute_risk(df, target_event, s, midpoint)
if mil_over:
df = preprocess_data_MIL_oversampling(df,target_event,mil_window_size,mil_thres)
if fs:
df = preprocess_data_feature_selection(df, target_event)
except Exception as err:
print(err)
traceback.print_exc(file=sys.stdout)
regr = RandomForestRegressor(n_jobs=rf_n_jobs, max_depth=rf_max_depth, random_state=rf_random_state,
n_estimators=rf_n_estimators)
train_df = df.iloc[:, :-3].reindex(sorted(df.iloc[:, :-3].columns), axis=1)
regr.fit(train_df, df['Risk'])
# print(regr.predict([df.iloc[0,:-1]]))
feature_importance = pd.DataFrame(regr.feature_importances_,
index=train_df.columns,
columns=['importance']).sort_values('importance', ascending=False)
return (regr, feature_importance)
def predict(regr, dataset, chunks_type, unique_event_ids=None):
df = preprocess_data_binarize(dataset, chunks_type, unique_event_ids, check_unknown_events=True)
return regr.predict(df.iloc[:, :-1])
# if (len(df) > 0):
# print(regr.predict(df))
# predictions.append(regr.predict(df))
# return predictions
# s = 0.7
# midpoint = 2
# target_event = 35544
# dataset = pd.read_csv('test-df.csv', parse_dates=True)
#
# train_rf_model(dataset, 'H', target_event, s, midpoint)
|