percent_change fails on top of aggregation primitives

[enfeizhan]

[A clear and concise description of what the bug is.]
The transform primitive percent_change fails to work on top of an aggregation primitive, returning NaN or seemingly arbitrary results.

Code Sample, a copy-pastable example to reproduce your bug.

These are the transactions:

transaction_id | customer_id | quantity | transaction_date
101758183 | abc | 15 | 2021-12-15
101862984 | abc | 15 | 2022-01-15
101960142 | abc | 15 | 2022-02-15
102062271 | abc | 15 | 2022-03-15
102179828 | abc | 15 | 2022-04-15
102301689 | abc | 15 | 2022-05-15
102434267 | abc | 15 | 2022-06-15
102540706 | abc | 15 | 2022-07-15
102662863 | abc | 15 | 2022-08-15
102783888 | abc | 15 | 2022-09-15
102901638 | abc | 15 | 2022-10-15
103041277 | abc | 15 | 2022-11-15
103199236 | abc | 15 | 2022-12-15
103336795 | abc | 15 | 2023-01-15
103478291 | abc | 15 | 2023-02-15
103604244 | abc | 15 | 2023-03-15
103738142 | abc | 15 | 2023-04-15
103895757 | abc | 15 | 2023-05-15
104073119 | abc | 15 | 2023-06-15
104233610 | abc | 15 | 2023-07-15

Creating the lables:
def is_churned(df):
return len(df) == 0

label_maker = cp.LabelMaker(
target_dataframe_index='customer_id',
time_index='transaction_date',
labeling_function=is_churned,
window_size='60d'
)

labels = label_maker.search(
df=tt1,
num_examples_per_instance=-1,
gap='1MS',
drop_empty=False,
minimum_data=6,
verbose=True
)
Labels will be like:
customer_id time is_churned
abc 2022-06-15 False
abc 2022-07-01 False
abc 2022-08-01 False
abc 2022-09-01 False
abc 2022-10-01 False
abc 2022-11-01 False
abc 2022-12-01 False
abc 2023-01-01 False
abc 2023-02-01 False
abc 2023-03-01 False
abc 2023-04-01 False
abc 2023-05-01 False
abc 2023-06-01 False
abc 2023-07-01 False

Create the EntitySet:
es = ft.EntitySet('bug')

es.add_dataframe(
dataframe=tt1,
dataframe_name='transactions',
time_index='transaction_date',
index='transaction_id'
)

es.normalize_dataframe(
base_dataframe_name='transactions',
new_dataframe_name='persons',
index='customer_id',
make_time_index=True
)

es.add_last_time_indexes()

Creating the features:
fm, fd = ft.dfs(
entityset=es,
target_dataframe_name='persons',
agg_primitives=['count'],
trans_primitives=['percent_change'],
cutoff_time=labels,
max_depth=2,
cutoff_time_in_index=True,
include_cutoff_time=False,
verbose=True,
)

customer_id time COUNT(transactions) PERCENT_CHANGE(COUNT(transactions)) is_churned
abc 2022-06-15 6 NaN False
abc 2022-07-01 7 NaN False
abc 2022-08-01 8 NaN False
abc 2022-09-01 9 NaN False
abc 2022-10-01 10 NaN False
abc 2022-11-01 11 NaN False
abc 2022-12-01 12 NaN False
abc 2023-01-01 13 NaN False
abc 2023-02-01 14 NaN False
abc 2023-03-01 15 NaN False
abc 2023-04-01 16 NaN False
abc 2023-05-01 17 NaN False
abc 2023-06-01 18 NaN False
abc 2023-07-01 19 NaN False
Notice the feature PERCENT_CHANGE(COUNT(transactions)) has all NaN, which should be 0 for the first row and a value roughly 0.05 ~ 0.2.

I also noticed the result can be quite random in a large real dataset, which is hard to be reproduced here.

# Your code here

Output of featuretools.show_info()

[paste the output of featuretools.show_info() here below this line]
Featuretools version: 1.28.0
Featuretools installation directory: /Users/feizhan/Installs/miniconda3/envs/generic/lib/python3.9/site-packages/featuretools

SYSTEM INFO

python: 3.9.13.final.0
python-bits: 64
OS: Darwin
OS-release: 23.0.0
machine: x86_64
processor: i386
byteorder: little
LC_ALL: None
LANG: en_AU.UTF-8
LOCALE: en_AU.UTF-8

INSTALLED VERSIONS

numpy: 1.23.4
pandas: 1.5.1
tqdm: 4.65.0
cloudpickle: 2.2.1
dask: 2023.3.2
distributed: 2023.3.2
psutil: 5.9.3
pip: 22.3
setuptools: 65.5.0