def test_generate_rules_apriori_large():
"""
Test with lots of data.
This test will fail if the second argument to `_ap_genrules` is not
validated as non-empty before the recursive function call. We must have
if H_m_copy:
yield from _ap_genrules
for this test to pass.
"""
transactions = generate_transactions(num_transactions=100,
unique_items=30,
items_row=(1, 20),
seed=123)
itemsets, num_transactions = itemsets_from_transactions(transactions, 0.1)
min_conf = 0.3
rules_apri = generate_rules_apriori(itemsets, min_conf, num_transactions)
rules_naive = generate_rules_naively(itemsets, min_conf, num_transactions)
rules_apri = list(rules_apri)
rules_naive = list(rules_naive)
# Test equal length, since no duplicates should be returned by apriori
assert len(rules_apri) == len(rules_naive)
# Test equal results
assert set(rules_apri) == set(rules_naive)
def apriori2(
transactions: typing.Union[typing.List[tuple], typing.Callable],
min_support: float = 0.5,
min_confidence: float = 0.5,
max_length: int = 8,
verbosity: int = 0,
output_transaction_ids: bool = False,
):
itemsets, num_trans = itemsets_from_transactions(
transactions,
min_support,
max_length,
verbosity,
output_transaction_ids,
)
if itemsets and isinstance(next(iter(itemsets[1].values())), ItemsetCount):
itemsets_for_rules = _convert_to_counts(itemsets)
else:
itemsets_for_rules = itemsets
for itemset in itemsets.items():
for itemset2 in itemset[1]:
tmp = itemset[1][itemset2] / num_trans
if tmp > 0.7 and tmp < 0.75:
itemset[1][itemset2] = "[0.7 - 0.75]"
elif tmp < 0.8:
itemset[1][itemset2] = "[0.75 - 0.8]"
elif tmp < 0.85:
itemset[1][itemset2] = "[0.8 - 0.85]"
elif tmp < 0.90:
itemset[1][itemset2] = "[0.85 - 0.9]"
else:
itemset[1][itemset2] = "[0.9 - 1]"
return itemsets, []
def speeds():
"""
Test the naive rule finder vs. the simple one from the paper.
"""
import random
random.seed(123456)
transactions = generate_transactions(num_transactions=random.randint(
250, 500),
unique_items=random.randint(8, 9),
items_row=(10, 50))
itemsets, num_transactions = itemsets_from_transactions(transactions, 0.1)
import time
min_conf = 0.5
print(itemsets)
st = time.perf_counter()
rules_apri = generate_rules_apriori(itemsets, min_conf, num_transactions)
rules_apri = list(rules_apri)
time_formatted = round(time.perf_counter() - st, 40)
print('Fast apriori ran in {} s'.format(time_formatted))
st = time.perf_counter()
rules_simple = generate_rules_simple(itemsets, min_conf, num_transactions)
rules_simple = list(rules_simple)
time_formatted = round(time.perf_counter() - st, 40)
print('Simple apriori ran in {} s'.format(time_formatted))
st = time.perf_counter()
rules_naive = generate_rules_naively(itemsets, min_conf, num_transactions)
rules_naive = list(rules_naive)
time_formatted = round(time.perf_counter() - st, 40)
print('Naive apriori ran in {} s'.format(time_formatted))
def test_itemsets_from_a_file():
"""
Test generator feature.
"""
def file_generator(filename_):
"""
A file generator for testing.
"""
def generate_from_file():
with open(filename_) as file:
for line in file:
transactions = tuple(line.strip("\n").split(","))
yield transactions
return generate_from_file
base, filename = os.path.split(__file__)
gen_obj = file_generator(os.path.join(base, "transactions.txt"))
result, _ = itemsets_from_transactions(
gen_obj, min_support=4 / 4, output_transaction_ids=True
)
assert result[2] == {
("A", "C"): ItemsetCount(itemset_count=4, members={0, 1, 2, 3})
}
def test_itemsets_max_length(transactions, min_support):
"""
The that nothing larger than max length is returned.
"""
max_len = random.randint(1, 5)
result, _ = itemsets_from_transactions(list(transactions), min_support,
max_length=max_len)
assert all(list(k <= max_len for k in result.keys()))
def test_itemsets_from_transactions_stochastic(transactions, min_support):
"""
Test 50 random inputs.
"""
result, _ = itemsets_from_transactions(list(transactions), min_support)
naive_result, _ = itemsets_from_transactions_naive(list(transactions),
min_support)
for key in set.union(set(result.keys()), set(naive_result.keys())):
assert result[key] == naive_result[key]
def test_generate_rules_simple_vs_naive(transactions):
"""
Test the naive rule finder vs. the simple one from the paper.
"""
itemsets, num_transactions = itemsets_from_transactions(transactions, 0.25)
min_conf = 0.1
rules_naive = generate_rules_naively(itemsets, min_conf, num_transactions)
rules_simple = generate_rules_simple(itemsets, min_conf, num_transactions)
assert set(rules_naive) == set(rules_simple)
def apriori(
transactions: typing.List[tuple],
min_support: float = 0.5,
min_confidence: float = 0.5,
max_length: int = 8,
verbosity: int = 0,
):
"""
The classic apriori algorithm as described in 1994 by Agrawal et al.
The Apriori algorithm works in two phases. Phase 1 iterates over the
transactions several times to build up itemsets of the desired support
level. Phase 2 builds association rules of the desired confidence given the
itemsets found in Phase 1. Both of these phases may be correctly
implemented by exhausting the search space, i.e. generating every possible
itemset and checking it's support. The Apriori prunes the search space
efficiently by deciding apriori if an itemset possibly has the desired
support, before iterating over the entire dataset and checking.
Parameters
----------
transactions : list of tuples, or a callable returning a generator
The transactions may be either a list of tuples, where the tuples must
contain hashable items. Alternatively, a callable returning a generator
may be passed. A generator is not sufficient, since the algorithm will
exhaust it, and it needs to iterate over it several times. Therefore,
a callable returning a generator must be passed.
min_support : float
The minimum support of the rules returned. The support is frequency of
which the items in the rule appear together in the data set.
min_confidence : float
The minimum confidence of the rules returned. Given a rule X -> Y, the
confidence is the probability of Y, given X, i.e. P(Y|X) = conf(X -> Y)
max_length : int
The maximum length of the itemsets and the rules.
verbosity : int
The level of detail printing when the algorithm runs. Either 0, 1 or 2.
Examples
--------
>>> transactions = [('a', 'b', 'c'), ('a', 'b', 'd'), ('f', 'b', 'g')]
>>> itemsets, rules = apriori(transactions, min_confidence=1)
>>> rules
[{a} -> {b}]
"""
itemsets, num_trans = itemsets_from_transactions(
transactions, min_support, max_length, verbosity
)
rules = generate_rules_apriori(
itemsets, min_confidence, num_trans, verbosity
)
return itemsets, list(rules)
def test_itemsets_from_a_generator_callable():
"""
Test generator feature.
"""
def generator():
"""
A generator for testing.
"""
for i in range(4):
yield [j + i for j in range(5)]
itemsets, _ = itemsets_from_transactions(generator, min_support=3 / 4)
assert itemsets[3] == {(2, 3, 4): 3, (3, 4, 5): 3}
def test_itemsets_max_length(transactions, min_support):
"""
The that nothing larger than max length is returned.
"""
max_len = random.randint(1, 5)
result, _ = itemsets_from_transactions(
list(transactions),
min_support,
max_length=max_len,
output_transaction_ids=True,
)
assert all(list(k <= max_len for k in result.keys()))
for length, itemsets in result.items():
for itemset_count in itemsets.values():
assert all(isinstance(i, int) for i in itemset_count.members)
def test_generate_rules_naive_vs_apriori(transactions):
"""
Test the naive rule finder vs. the simple one from the paper.
"""
itemsets, num_transactions = itemsets_from_transactions(transactions, 0.15)
min_conf = 0.3
rules_apri = generate_rules_apriori(itemsets, min_conf, num_transactions)
rules_naive = generate_rules_naively(itemsets, min_conf, num_transactions)
rules_apri = list(rules_apri)
rules_naive = list(rules_naive)
# Test equal length, since no duplicates should be returned by apriori
assert len(rules_apri) == len(rules_naive)
# Test equal results
assert set(rules_apri) == set(rules_naive)
def test_itemsets_from_a_file():
"""
Test generator feature.
"""
def file_generator(filename):
"""
A file generator for testing.
"""
def generate_from_file():
with open(filename) as file:
for line in file:
yield tuple(line.strip("\n").split(","))
return generate_from_file
base, filename = os.path.split(__file__)
gen_obj = file_generator(os.path.join(base, "transactions.txt"))
result, _ = itemsets_from_transactions(gen_obj, min_support=4 / 4)
assert result[2] == {("A", "C"): 4}
def test_itemsets_from_a_generator_callable():
"""
Test generator feature.
"""
def generator():
"""
A generator for testing.
"""
for i in range(4):
transactions = tuple(j + i for j in range(5))
yield transactions
itemsets, _ = itemsets_from_transactions(generator,
min_support=3 / 4,
output_transaction_ids=True)
assert itemsets[3] == {
(2, 3, 4): ItemsetCount(itemset_count=3, members={0, 1, 2}),
(3, 4, 5): ItemsetCount(itemset_count=3, members={1, 2, 3}),
}
def apriori(
transactions: typing.Iterable[typing.Union[set, tuple, list]],
min_support: float = 0.5,
min_confidence: float = 0.5,
max_length: int = 8,
verbosity: int = 0,
output_transaction_ids: bool = False,
):
"""
The classic apriori algorithm as described in 1994 by Agrawal et al.
The Apriori algorithm works in two phases. Phase 1 iterates over the
transactions several times to build up itemsets of the desired support
level. Phase 2 builds association rules of the desired confidence given the
itemsets found in Phase 1. Both of these phases may be correctly
implemented by exhausting the search space, i.e. generating every possible
itemset and checking it's support. The Apriori prunes the search space
efficiently by deciding apriori if an itemset possibly has the desired
support, before iterating over the entire dataset and checking.
Parameters
----------
transactions : list of transactions (sets/tuples/lists). Each element in
the transactions must be hashable.
min_support : float
The minimum support of the rules returned. The support is frequency of
which the items in the rule appear together in the data set.
min_confidence : float
The minimum confidence of the rules returned. Given a rule X -> Y, the
confidence is the probability of Y, given X, i.e. P(Y|X) = conf(X -> Y)
max_length : int
The maximum length of the itemsets and the rules.
verbosity : int
The level of detail printing when the algorithm runs. Either 0, 1 or 2.
output_transaction_ids : bool
If set to true, the output contains the ids of transactions that
contain a frequent itemset. The ids are the enumeration of the
transactions in the sequence they appear.
Examples
--------
>>> transactions = [('a', 'b', 'c'), ('a', 'b', 'd'), ('f', 'b', 'g')]
>>> itemsets, rules = apriori(transactions, min_confidence=1)
>>> rules
[{a} -> {b}]
"""
itemsets, num_trans = itemsets_from_transactions(
transactions,
min_support,
max_length,
verbosity,
output_transaction_ids=True,
)
itemsets_raw = {
length:
{item: counter.itemset_count
for (item, counter) in itemsets.items()}
for (length, itemsets) in itemsets.items()
}
rules = generate_rules_apriori(itemsets_raw, min_confidence, num_trans,
verbosity)
if output_transaction_ids:
return itemsets, list(rules)
else:
return itemsets_raw, list(rules)
def apriori(
transactions,
min_support=0.5,
min_confidence=0.5,
max_length=8,
verbosity=0,
output_transaction_ids=False,
):
"""
The classic apriori algorithm as described in 1994 by Agrawal et al.
The Apriori algorithm works in two phases. Phase 1 iterates over the
transactions several times to build up itemsets of the desired support
level. Phase 2 builds association rules of the desired confidence given the
itemsets found in Phase 1. Both of these phases may be correctly
implemented by exhausting the search space, i.e. generating every possible
itemset and checking it's support. The Apriori prunes the search space
efficiently by deciding apriori if an itemset possibly has the desired
support, before iterating over the entire dataset and checking.
Parameters
----------
transactions : list of tuples, list of itemsets.TransactionWithId,
or a callable returning a generator. Use TransactionWithId's when
the transactions have ids which should appear in the outputs.
The transactions may be either a list of tuples, where the tuples must
contain hashable items. Alternatively, a callable returning a generator
may be passed. A generator is not sufficient, since the algorithm will
exhaust it, and it needs to iterate over it several times. Therefore,
a callable returning a generator must be passed.
min_support : float
The minimum support of the rules returned. The support is frequency of
which the items in the rule appear together in the data set.
min_confidence : float
The minimum confidence of the rules returned. Given a rule X -> Y, the
confidence is the probability of Y, given X, i.e. P(Y|X) = conf(X -> Y)
max_length : int
The maximum length of the itemsets and the rules.
verbosity : int
The level of detail printing when the algorithm runs. Either 0, 1 or 2.
output_transaction_ids : bool
If set to true, the output contains the ids of transactions that
contain a frequent itemset. The ids are the enumeration of the
transactions in the sequence they appear.
Examples
--------
>>> transactions = [('a', 'b', 'c'), ('a', 'b', 'd'), ('f', 'b', 'g')]
>>> itemsets, rules = apriori(transactions, min_confidence=1)
>>> rules
[{a} -> {b}]
"""
print('itemsets_from_transactions')
itemsets, num_trans = itemsets_from_transactions(
transactions,
min_support,
max_length,
verbosity,
output_transaction_ids,
)
if itemsets and isinstance(next(iter(itemsets[1].values())), ItemsetCount):
itemsets_for_rules = _convert_to_counts(itemsets)
else:
itemsets_for_rules = itemsets
print('generate_rules_apriori')
rules = generate_rules_apriori(itemsets_for_rules, min_confidence,
num_trans, verbosity)
return itemsets, list(rules)
