def _calculate_rule_confidence(self, car, pandas_df):
quant_dataframe = QuantitativeDataFrame(pandas_df)
support, confidence = quant_dataframe.calculate_rule_statistics(car)
return confidence
def fit(self, rule_cutoff):
dataframes = self._prepare_dataframes()
scores = []
for dataframe_train, dataframe_test in dataframes:
txns_train = TransactionDB.from_DataFrame(dataframe_train)
rules = top_rules(txns_train.string_representation, appearance=txns_train.appeardict)
cars = createCARs(rules)[:rule_cutoff]
quant_dataframe_train = QuantitativeDataFrame(dataframe_train)
quant_dataframe_test = QuantitativeDataFrame(dataframe_test)
self.classifier.fit(quant_dataframe_train, cars, debug=self.debug)
score = None
if self.score_auc:
score = self.classifier.score_auc(quant_dataframe_test)
else:
score = self.classifier.score(quant_dataframe_test)
scores.append(score)
return scores
def _prepare_data_sample(self, quant_dataframe):
pandas_dataframe = quant_dataframe.dataframe
class_column = pandas_dataframe[
self.class_name] if self.class_name else pandas_dataframe.iloc[:,
-1]
unique_classes = np.unique(class_column.values)
restricted_quant_dataframes = []
for class_ in unique_classes:
# TODO
# find a better way than copy
dataframe_restricted = pandas_dataframe.copy()
dataframe_class_column_restricted = np.where(
class_column == class_, class_, self.other_class_label)
if self.class_name:
dataframe_restricted[
self.class_name] = dataframe_class_column_restricted
else:
dataframe_restricted.iloc[:,
-1] = dataframe_class_column_restricted
dataframe = QuantitativeDataFrame(dataframe_restricted)
restricted_quant_dataframes.append(dataframe)
return restricted_quant_dataframes
def split_data_by_class(self, quant_dataframe):
pandas_dataframe = quant_dataframe.dataframe
class_column = pandas_dataframe[
self.class_name] if self.class_name else pandas_dataframe.iloc[:,
-1]
unique_classes = np.unique(class_column.values)
restricted_quant_dataframes = dict()
for class_ in unique_classes:
dataframe_restricted = pandas_dataframe.copy()
dataframe_class_column_restricted = np.where(
class_column == class_, class_, self.other_class_label)
if self.class_name:
dataframe_restricted[
self.class_name] = dataframe_class_column_restricted
else:
dataframe_restricted.iloc[:,
-1] = dataframe_class_column_restricted
dataframe = QuantitativeDataFrame(dataframe_restricted)
restricted_quant_dataframes[class_] = dataframe
return restricted_quant_dataframes
def test_optimization(self):
df = pd.read_csv("C:/code/python/interpretable_decision_sets/data/titanic.csv")
ids_ruleset = mine_IDS_ruleset(df, rule_cutoff=40)
quant_df = QuantitativeDataFrame(df)
ascent = CoordinateAscentOptimizer(IDS(), maximum_consecutive_iterations=1)
lambdas = ascent.fit(ids_ruleset, quant_df, quant_df)
def _prepare(self, quant_dataframe, class_name):
if type(quant_dataframe) != QuantitativeDataFrame:
raise Exception("Type of quant_dataframe must be QuantitativeDataFrame")
self.quant_dataframe = quant_dataframe
self.pandas_dataframe = self.quant_dataframe.dataframe
self.ids_classifiers = dict()
self.class_name = class_name
self.other_class_label = "OTHER"
class_column = self.pandas_dataframe[class_name] if class_name else self.pandas_dataframe.iloc[:,-1]
unique_classes = np.unique(class_column.values)
if len(unique_classes) < 3:
raise Exception("Number of distinct classes must be greater than 2, otherwise use binary classifier")
for class_ in unique_classes:
# TODO
# find a better way than copy
dataframe_restricted = self.pandas_dataframe.copy()
dataframe_class_column_restricted = np.where(class_column == class_, class_, self.other_class_label)
if class_name:
dataframe_restricted[class_name] = dataframe_class_column_restricted
else:
dataframe_restricted.iloc[:,-1] = dataframe_class_column_restricted
ids_class_clf = IDS()
self.ids_classifiers.update({class_ : dict(
quant_dataframe=QuantitativeDataFrame(dataframe_restricted),
clf=ids_class_clf
)})
def test_model_fitting(self):
df = pd.read_csv(
"C:/code/python/interpretable_decision_sets/data/iris0.csv")
quant_df = QuantitativeDataFrame(df)
ids = IDSOneVsAll()
ids.fit(quant_df, quant_df, debug=False)
auc = ids.score_auc(quant_df)
def test_dls_algorithm(self):
df = pd.read_csv(
"C:/code/python/interpretable_decision_sets/data/iris0.csv")
cars = mine_CARs(df, rule_cutoff=40)
quant_df = QuantitativeDataFrame(df)
ids = IDSOneVsAll()
ids.fit(quant_df, quant_df, algorithm="SLS", debug=False)
auc = ids.score_auc(quant_df)
def test_model_fitting(self):
df = pd.read_csv(
"C:/code/python/interpretable_decision_sets/data/titanic.csv")
cars = mine_CARs(df, rule_cutoff=40)
quant_df = QuantitativeDataFrame(df)
ids = IDS()
ids.fit(quant_df, cars, debug=False)
auc = ids.score_auc(quant_df)
def test_random_seed(self):
replications_n = 10
cars_to_mine = 10
df = pd.read_csv("data/iris0.csv")
quant_df = QuantitativeDataFrame(df)
mined_cars_mupliple = []
mined_cars_comparison_results = []
for _ in range(replications_n):
ClassAssocationRule.id = 0
cars = mine_CARs(df, cars_to_mine)
mined_cars_mupliple.append(cars)
for idx in range(replications_n):
same = _all_rules_same(mined_cars_mupliple[0],
mined_cars_mupliple[idx])
mined_cars_comparison_results.append(same)
self.assertTrue(np.all(mined_cars_comparison_results))
ids_models_multiple = []
ids_comparison_results = []
for _ in range(replications_n):
ids = IDS()
ids = ids.fit(quant_dataframe=quant_df,
class_association_rules=cars,
debug=False,
random_seed=2)
ids_models_multiple.append(ids.clf.rules)
for idx in range(replications_n):
same = _all_rules_same(ids_models_multiple[0],
ids_models_multiple[idx])
ids_comparison_results.append(same)
self.assertTrue(np.all(ids_comparison_results))
def _prepare(self, quant_dataframe: QuantitativeDataFrame,
class_name: str):
if type(quant_dataframe) != QuantitativeDataFrame:
raise Exception(
"Type of quant_dataframe must be QuantitativeDataFrame")
self.quant_dataframe = quant_dataframe
self.pandas_dataframe = self.quant_dataframe.dataframe
self.class_name = class_name if class_name else self.pandas_dataframe.columns[
-1]
class_column = self.pandas_dataframe[self.class_name]
unique_classes = np.unique(class_column.values)
if len(unique_classes) < 3:
raise Exception(
"Number of distinct classes must be greater than 2, otherwise use binary classifier"
)
for class_ in unique_classes:
dataframe_restricted = self.pandas_dataframe.copy()
dataframe_class_column_restricted = np.where(
class_column == class_, class_, self.other_class_label)
dataframe_restricted[
self.class_name] = dataframe_class_column_restricted
ids_class_clf = IDS(algorithm=self.algorithm)
self.ids_classifiers.update({
class_:
dict(quant_dataframe=QuantitativeDataFrame(
dataframe_restricted),
rules=None,
clf=ids_class_clf)
})
from pyids.algorithms import mine_CARs
from pyids.data_structures import IDSRuleSet
import json
from pyarc.qcba.data_structures import QuantitativeDataFrame
import random
import logging
import time
logging.basicConfig(level=logging.INFO)
df = pd.read_csv("../../../data/iris0.csv")
cars = mine_CARs(df, 10, sample=False)
ids_ruleset = IDSRuleSet.from_cba_rules(cars).ruleset
quant_dataframe = QuantitativeDataFrame(df)
def generate_lambda_arr(one_idx):
total_params = 7
if one_idx == 0:
start_arr = []
else:
start_arr = one_idx * [0]
end_arr = (total_params - one_idx - 1) * [0]
return start_arr + [1] + end_arr
import pandas as pd
from pyids.ids_classifier import IDSOneVsAll, mine_IDS_ruleset
from pyids.model_selection import CoordinateAscentOptimizer, train_test_split_pd
from pyarc.qcba.data_structures import QuantitativeDataFrame
df = pd.read_csv("./data/iris0.csv")
df_train, df_test = train_test_split_pd(df, prop=0.2)
ids_ruleset = mine_IDS_ruleset(df_train, rule_cutoff=50)
quant_dataframe_train = QuantitativeDataFrame(df_train)
quant_dataframe_test = QuantitativeDataFrame(df_test)
coordinate_ascent = CoordinateAscentOptimizer(
IDSOneVsAll(),
debug=True,
maximum_delta_between_iterations=200,
maximum_score_estimation_iterations=3)
coordinate_ascent.fit(ids_ruleset, quant_dataframe_train, quant_dataframe_test)
best_lambda_array = coordinate_ascent.current_best_params
from pyids.ids_classifier import IDS, mine_IDS_ruleset
from pyids.ids_ruleset import IDSRuleSet
from pyids.model_selection import RandomSearchOptimizer, train_test_split_pd
from pyids.rule_mining import RuleMiner
df = pd.read_csv("../../data/titanic.csv")
rm = RuleMiner()
cars = rm.mine_rules(df, minsup=0.005)
ids_ruleset = IDSRuleSet.from_cba_rules(cars)
df_train, df_test = train_test_split_pd(df, prop=0.25)
quant_df_train, quant_df_test = QuantitativeDataFrame(df_train), QuantitativeDataFrame(df_test)
random_optimizer = RandomSearchOptimizer(IDS(), maximum_score_estimation_iterations=5, maximum_iterations=500)
lambda_array = random_optimizer.fit(ids_ruleset, quant_df_train, quant_df_test)
all_params = random_optimizer.score_params_dict
print(lambda_array)
with open("results/random_search_lambda_array.txt", "w") as file:
file.write(str(lambda_array))
with open("results/random_search_all_score_params.txt", "w") as file:
file.write(str(random_optimizer.score_params_dict))
return [x for x in os.listdir(path) if x.startswith(dataset_name)]
benchmark_data = []
for dataset_name, rule_cutoff in dataset_rulenum.items():
print(dataset_name)
train_files = get_dataset_files(train_path, dataset_name)
test_files = get_dataset_files(test_path, dataset_name)
for train_file, test_file in list(zip(train_files, test_files))[:]:
dataset_path = os.path.join(train_path, train_file)
dataset_test_path = os.path.join(test_path, test_file)
df = pd.read_csv(dataset_path)
quant_df = QuantitativeDataFrame(df)
txns = TransactionDB.from_DataFrame(df)
df_test = pd.read_csv(dataset_test_path)
quant_df_test = QuantitativeDataFrame(df_test)
txns_test = TransactionDB.from_DataFrame(df_test)
def fmax(param_dict):
print(param_dict)
support, confidence = param_dict["support"] / 1000, param_dict[
"confidence"] / 1000
print(dict(support=support, confidence=confidence))
cba = CBA(support=support, confidence=confidence)
cba.fit(txns)
import pandas as pd
import time
import numpy as np
import random
from pyarc import CBA
from pyarc.qcba import QCBA
from pyarc import TransactionDB
from pyarc.algorithms import (top_rules, createCARs, M1Algorithm, M2Algorithm)
df = pd.read_csv(
"c:/code/python/machine_learning/assoc_rules/train/lymph0.csv")
df_undiscr = pd.read_csv(
"c:/code/python/machine_learning/assoc_rules/folds_undiscr/train/lymph0.csv"
)
quant_df = QuantitativeDataFrame(df)
quant_df_undiscr = QuantitativeDataFrame(df_undiscr)
benchmark_data = []
time_estimation_iterations = 10
max_rules = 100
def generate_lambda_array():
lambdas = [generate_lambda_parameter() for i in range(7)]
return lambdas
def generate_lambda_parameter():
from pyids.algorithms.ids import IDS
from pyids.algorithms import mine_CARs
from pyarc.qcba.data_structures import QuantitativeDataFrame
from pyids.model_selection.coordinate_ascent import CoordinateAscent
lambda_dict = {
'l1': 124.16415180612711,
'l2': 38.896662094192955,
'l3': 557.0996799268405,
'l4': 638.188385916781,
'l5': 136.48056698673983,
'l6': 432.1760402377687,
'l7': 452.1563786008231
}
lambda_array = [
665.9341563786008, 271.7242798353909, 212.34156378600824,
20.489711934156375, 648.5761316872428, 911, 560
]
df = pd.read_csv("C:/code/python/machine_learning/assoc_rules/train/iris0.csv")
quant_df = QuantitativeDataFrame(df)
quant_df_test = QuantitativeDataFrame(
pd.read_csv("C:/code/python/machine_learning/assoc_rules/test/iris0.csv"))
cars = mine_CARs(df, 20)
ids = IDS(algorithm="DUSM")
ids.fit(quant_df, cars, lambda_array=lambda_array)
print(ids.score_auc(quant_df))
def generate_lambda_array():
lambdas = [generate_lambda_parameter() for i in range(7)]
return lambdas
def generate_lambda_parameter():
return random.randint(0, 1000)
for i in range(100, 3000, 100):
df = df_all.iloc[:i, :].copy()
data_count = df.shape[0]
quant_df = QuantitativeDataFrame(df)
for algorithm in ["DLS", "SLS", "DUSM", "RUSM"]:
durations = []
for _ in range(time_estimation_iterations):
print(_)
lambda_array = generate_lambda_array()
print(f"data count: {data_count}")
print(f"algorithm: {algorithm}")
print(f"using lambda: {lambda_array}")
cars = mine_CARs(df, rule_cutoff=rule_cutoff)
ids = IDS(algorithm=algorithm)
def run1fold(basepath,
datasetname,
unique_transactions=True,
runQCBA=False,
saveIDSRules=True,
useConfidenceForCandidateGeneration=True):
df_stat = pd.DataFrame(
columns=['ids', 'idsqcba'],
index=["accuracy", "rulecount", "rulelength", "buildtime"])
if (runQCBA):
#python QCBA implementation uses custom discretization format
data_train_disc = pd.read_csv(
basepath + "data/folds_discr/train/{}.csv".format(datasetname))
data_test_disc = pd.read_csv(
basepath + "data/folds_discr/test/{}.csv".format(datasetname))
data_test_undisc = pd.read_csv(
basepath + "data/folds_nodiscr/test/{}.csv".format(datasetname))
data_train_undisc = pd.read_csv(
basepath + "data/folds_nodiscr/train/{}.csv".format(datasetname))
quant_dataframe_test_undisc = QuantitativeDataFrame(data_test_undisc)
quant_dataframe_train_undisc = QuantitativeDataFrame(data_train_undisc)
else:
#R QCBA implementation uses different discretization format, folds are generated with preprocess_for_ids.R
data_train_disc = pd.read_csv(
basepath + "data/folds_discr2/train/{}.csv".format(datasetname))
data_test_disc = pd.read_csv(
basepath + "data/folds_discr2/test/{}.csv".format(datasetname))
quant_dataframe_train_disc = QuantitativeDataFrame(data_train_disc)
quant_dataframe_test_disc = QuantitativeDataFrame(data_test_disc)
actual = quant_dataframe_test_disc.dataframe.iloc[:, -1].values
if useConfidenceForCandidateGeneration:
# mine_CARs learns initial candidate rules with CBA-like approach
# it uses unsupervised paramter tuning to determine conf, supp and len thresholds,
# as described in Kliegr & Kuchar, 2019
# Because the subsequent optimization is slow, not all initial candidate rules can be passed to IDS.
# the sample parameter controls, how the subset of N rules will be selected from the initial candidates:
# sample=False: take top N rules according to CBA criteria. According to our experiments, this has better results
# sample=True: take random N rules
cars = mine_CARs(data_train_disc, 50, sample=False)
else:
# learn candidate rules using approach without min confidence described in Lakkaraju et al, 2-16
print("WARNING save any unsaved work")
print(
"WARNING candidate generation without minimum confidence and sampling may be too slow or memory intensive"
)
rm = RuleMiner()
cars = rm.mine_rules(
data_train_disc,
minsup=0.01) # the 0.01 threshold is from the IDS paper
print(len(cars))
print("rule mining finished")
#train IDS model
ids = IDS()
start = time.time()
# all lambdas are set to the same value
ids.fit(class_association_rules=cars,
lambda_array=7 * [1],
quant_dataframe=quant_dataframe_train_disc,
debug=False,
random_seed=1)
end = time.time()
df_stat.loc["buildtime", "ids"] = end - start
#apply IDS model
df_stat.loc["accuracy", "ids"] = ids.score(quant_dataframe_test_disc)
print("Acc IDS:", df_stat.loc["accuracy", "ids"])
df_stat.loc["rulecount", "ids"] = len(ids.clf.rules)
antLengths = list(
map(lambda r: len(r.car.antecedent.itemset.items()), ids.clf.rules))
df_stat.loc["rulelength", "ids"] = sum(antLengths) / len(antLengths)
avg_rule_legnth_ids = None
print("Rule Count IDS:", df_stat.loc["rulecount", "ids"])
if (saveIDSRules):
idsRulesPath = basepath + modelFolder + "/{}.csv".format(datasetname)
file = open(idsRulesPath, "w")
txtexport = "rules,suppport,confidence,lift\n"
#Before export, IDS sorts the rule by harmonic mean of support and confidence (st.hmean([self.car.support, self.car.confidence]))
#In this order, rules are also applied for prediction
for r in ids.clf.rules:
args = [
r.car.antecedent.string(),
"{" + r.car.consequent.string() + "}", r.car.support,
r.car.confidence, 0
]
txtexport = txtexport + "\"{} => {}\",{:.2f},{:.2f},{:.2f} \n".format(
*args)
#add default rule
classname = data_train_disc.columns.values[-1]
txtexport = txtexport + "\"{ } => " + "{" + classname + "=" + mode(
data_train_disc[data_train_disc.columns[-1]]) + "}\", 0,0,0"
print(txtexport)
file.write(txtexport)
file.close()
if (runQCBA):
#postprocess IDS model with QCBA
rules_to_optimize = ids.clf.rules
start = time.time()
quant_rules = [QuantitativeCAR(r.car) for r in rules_to_optimize]
qcba_transformation = QCBATransformation(quant_dataframe_train_undisc)
transformed_rules = qcba_transformation.transform(quant_rules)
end = time.time()
df_stat.loc["buildtime", "idsqcba"] = end - start
rules, default_class = transformed_rules
antLengths = list(
map(lambda r: len(r.car.antecedent.itemset.items()),
ids.clf.rules))
#+1 because the default rule is not counted
df_stat.loc["rulelength",
"idsqcba"] = sum(antLengths) / (len(antLengths) + 1)
#apply QCBA model
qclf = QuantitativeClassifier(rules, default_class)
pred = qclf.predict(quant_dataframe_test_undisc)
#evaluate model - QCBA
df_stat.loc["accuracy", "idsqcba"] = accuracy_score(actual, pred)
df_stat.loc["rulecount", "idsqcba"] = len(rules)
print("Acc IDS-QCBA:", df_stat.loc["accuracy", "idsqcba"])
print("Rule Count IDS-QCBA:", df_stat.loc["rulecount", "idsqcba"])
return df_stat