prescriptiveProcessMonitoring/optimize_threshold_compensation.py at master · samadeusfp/prescriptiveProcessMonitoring · GitHub

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
import EncoderFactory
from DatasetManager import DatasetManager

import pandas as pd
import numpy as np

from sklearn.metrics import roc_auc_score
from sklearn.pipeline import FeatureUnion

import time
import os
import sys
from sys import argv
import pickle
import csv

from conf_constant_costfunctions import get_constant_costfunctions

from hyperopt import Trials, STATUS_OK, tpe, fmin, hp
import hyperopt
from multiprocessing import Process as Process


def calculate_cost(x, costs):
    return costs[int(x['prediction']), int(x['actual'])](x)

def evaluate_model_cost(args):
    conf_threshold = args['conf_threshold']
    c_action = args['c_action']
    c_miss = args['c_miss']
    c_com = args['c_com']

    if early_type == "linear":
        costs = np.matrix([[lambda x: 0,
                            lambda x: c_miss],
                           [lambda x: c_action * (x['prefix_nr'] - 1) / x['case_length'] + c_com,
                            lambda x: c_action * (x['prefix_nr'] - 1) / x['case_length'] + (x['prefix_nr']) / x[
                                'case_length'] * c_miss
                            ]])
    elif early_type == "nonmonotonic":
        costs = np.matrix([[lambda x: 0,
                            lambda x: c_miss],
                           [lambda x: (c_action * (
                                   1 - min(x['prefix_nr'], aConst) / bConst)) + (c_com * (
                                   1 - min(x['prefix_nr'], cConst) / dConst)),
                            lambda x: (c_action * (
                                   1 - min(x['prefix_nr'], aConst) / bConst)) + (1 - max(min(eConst,x['prefix_nr'])/fConst,1) *  c_miss)
                            ]])
    else:
        costs = np.matrix([[lambda x: 0,
                            lambda x: c_miss],
                           [lambda x: c_action + c_com,
                            lambda x: c_action + (x['prefix_nr'] - 1) / x['case_length'] * c_miss
                            ]])

    # trigger alarms according to conf_threshold
    dt_final = pd.DataFrame()
    unprocessed_case_ids = set(dt_preds.case_id.unique())
    for nr_events in range(1, dt_preds.prefix_nr.max() + 1):
        tmp = dt_preds[(dt_preds.case_id.isin(unprocessed_case_ids)) & (dt_preds.prefix_nr == nr_events)]
        tmp = tmp[tmp.predicted_proba >= conf_threshold]
        tmp["prediction"] = 1
        dt_final = pd.concat([dt_final, tmp], axis=0)
        unprocessed_case_ids = unprocessed_case_ids.difference(tmp.case_id)
    tmp = dt_preds[(dt_preds.case_id.isin(unprocessed_case_ids)) & (dt_preds.prefix_nr == 1)]
    tmp["prediction"] = 0
    dt_final = pd.concat([dt_final, tmp], axis=0)

    case_lengths = dt_preds.groupby("case_id").prefix_nr.max().reset_index()
    case_lengths.columns = ["case_id", "case_length"]
    dt_final = dt_final.merge(case_lengths)

    cost = dt_final.apply(calculate_cost, costs=costs, axis=1).sum()

    return {'loss': cost, 'status': STATUS_OK, 'model': dt_final}

def run_experiment(c_miss_weight,c_action_weight,c_com_weight,early_type):
    c_miss = c_miss_weight / (c_miss_weight + c_action_weight + c_com_weight)
    c_action = c_action_weight / (c_miss_weight + c_action_weight + c_com_weight)
    c_com = c_com_weight / (c_miss_weight + c_action_weight + c_com_weight)

    space = {'conf_threshold': hp.uniform("conf_threshold", 0, 1),
             'c_action': c_action,
             'c_miss': c_miss,
             'c_com': c_com}
    trials = Trials()
    best = fmin(evaluate_model_cost, space, algo=tpe.suggest, max_evals=50, trials=trials)

    best_params = hyperopt.space_eval(space, best)

    outfile = os.path.join(params_dir, "optimal_confs_%s_%s_%s_%s_%s_%s.pickle" % (
    dataset_name, c_miss_weight, c_action_weight, c_postpone_weight, c_com_weight, early_type))
    # write to file
    with open(outfile, "wb") as fout:
        pickle.dump(best_params, fout)


print('Preparing data...')
start = time.time()

dataset_name = argv[1]
preds_dir = argv[2]
params_dir = argv[3]


# create output directory
if not os.path.exists(os.path.join(params_dir)):
    os.makedirs(os.path.join(params_dir))

# read the data
dataset_manager = DatasetManager(dataset_name)

# prepare the dataset
dt_preds = pd.read_csv(os.path.join(preds_dir, "preds_val_%s.csv" % dataset_name), sep=";")

#set nonomonotic constants
aConst, bConst, cConst, dConst, eConst, fConst = get_constant_costfunctions(dataset_name)


print('Optimizing parameters...')
processes = []
cost_weights = [(10, 1), (10, 2), (10, 3), (10, 4), (10, 5)]
c_com_weights = [0,1,2,3,4,5,10,20]
c_postpone_weight = 0
for c_miss_weight, c_action_weight in cost_weights:
    for c_com_weight in c_com_weights:
        for early_type in ["const", "linear","nonmonotonic"]:
            p = Process(target=run_experiment, args=(c_miss_weight, c_action_weight, c_com_weight, early_type))
            p.start()
            processes.append(p)
for p in processes:
    p.join()