Source code for evalml.data_checks.datetime_format_data_check
"""Data check that checks if the datetime column has equally spaced intervals and is monotonically increasing or decreasing in order to be supported by time series estimators."""
import pandas as pd
from evalml.data_checks import DataCheck, DataCheckError, DataCheckMessageCode
from evalml.utils import infer_feature_types
[docs]class DateTimeFormatDataCheck(DataCheck):
"""Check if the datetime column has equally spaced intervals and is monotonically increasing or decreasing in order to be supported by time series estimators.
Args:
datetime_column (str, int): The name of the datetime column. If the datetime values are in the index, then pass "index".
"""
def __init__(self, datetime_column="index"):
self.datetime_column = datetime_column
[docs] def validate(self, X, y):
"""Checks if the target data has equal intervals and is sorted.
Args:
X (pd.DataFrame, np.ndarray): Features.
y (pd.Series, np.ndarray): Target data.
Returns:
dict (DataCheckError): List with DataCheckErrors if unequal intervals are found in the datetime column.
Examples:
>>> import pandas as pd
The column "dates" has a set of dates with hourly frequency appended to the end of a series of days, which is inconsistent
with the frequency of the previous 9 dates (1 day).
>>> X = pd.DataFrame(pd.date_range("2021-01-01", periods=6).append(pd.date_range("2021-01-07", periods=3, freq="H")), columns=["dates"])
>>> y = pd.Series([0, 1, 0, 1, 1, 0, 0, 0, 1, 0])
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="dates")
>>> assert datetime_format_dc.validate(X, y) == [
... {
... "message": "Column 'dates' has datetime values missing between start and end date.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "code": "DATETIME_IS_MISSING_VALUES",
... "details": {"columns": None, "rows": None},
... "action_options": []
... },
... {
... "message": "No frequency could be detected in column 'dates', possibly due to uneven intervals.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "code": "DATETIME_HAS_UNEVEN_INTERVALS",
... "details": {"columns": None, "rows": None},
... "action_options": []
... }
... ]
The column "dates" has the date 2021-01-31 appended to the end, which implies there are many dates missing.
>>> X = pd.DataFrame(pd.date_range("2021-01-01", periods=9).append(pd.date_range("2021-01-31", periods=1)), columns=["dates"])
>>> y = pd.Series([0, 1, 0, 1, 1, 0, 0, 0, 1, 0])
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="dates")
>>> assert datetime_format_dc.validate(X, y) == [
... {
... "message": "Column 'dates' has datetime values missing between start and end date.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "code": "DATETIME_IS_MISSING_VALUES",
... "details": {"columns": None, "rows": None},
... "action_options": []
... }
... ]
The column "dates" has a repeat of the date 2021-01-09 appended to the end, which is considered redundant and will raise an error.
>>> X = pd.DataFrame(pd.date_range("2021-01-01", periods=9).append(pd.date_range("2021-01-09", periods=1)), columns=["dates"])
>>> y = pd.Series([0, 1, 0, 1, 1, 0, 0, 0, 1, 0])
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="dates")
>>> assert datetime_format_dc.validate(X, y) == [
... {
... "message": "Column 'dates' has more than one row with the same datetime value.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "code": "DATETIME_HAS_REDUNDANT_ROW",
... "details": {"columns": None, "rows": None},
... "action_options": []
... }
... ]
The column "Weeks" passed integers instead of datetime data, which will raise an error.
>>> X = pd.DataFrame([1, 2, 3, 4], columns=["Weeks"])
>>> y = pd.Series([0] * 4)
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="Weeks")
>>> assert datetime_format_dc.validate(X, y) == [
... {
... "message": "Datetime information could not be found in the data, or was not in a supported datetime format.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "details": {"columns": None, "rows": None},
... "code": "DATETIME_INFORMATION_NOT_FOUND",
... "action_options": []
... }
... ]
Converting that same integer data to datetime, however, is valid.
>>> X = pd.DataFrame(pd.to_datetime([1, 2, 3, 4]), columns=["Weeks"])
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="Weeks")
>>> assert datetime_format_dc.validate(X, y) == []
>>> X = pd.DataFrame(pd.date_range("2021-01-01", freq="W", periods=10), columns=["Weeks"])
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="Weeks")
>>> assert datetime_format_dc.validate(X, y) == []
While the data passed in is of datetime type, time series requires the datetime information in datetime_column
to be monotonically increasing (ascending).
>>> X = X.iloc[::-1]
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="Weeks")
>>> assert datetime_format_dc.validate(X, y) == [
... {
... "message": "Datetime values must be sorted in ascending order.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "details": {"columns": None, "rows": None},
... "code": "DATETIME_IS_NOT_MONOTONIC",
... "action_options": []
... }
... ]
The first value in the column "index" is replaced with NaT, which will raise an error in this data check.
>>> dates = [["2-1-21", "3-1-21"],
... ["2-2-21", "3-2-21"],
... ["2-3-21", "3-3-21"],
... ["2-4-21", "3-4-21"]]
>>> dates[0][0] = None
>>> df = pd.DataFrame(dates, columns=["days", "days2"])
>>> datetime_format_dc = DateTimeFormatDataCheck(datetime_column="days")
>>> assert datetime_format_dc.validate(df, y) == [
... {
... "message": "Input datetime column 'days' contains NaN values. Please impute NaN values or drop these rows.",
... "data_check_name": "DateTimeFormatDataCheck",
... "level": "error",
... "details": {"columns": None, "rows": None},
... "code": "DATETIME_HAS_NAN",
... "action_options": []
... }
... ]
...
"""
messages = []
X = infer_feature_types(X)
y = infer_feature_types(y)
no_dt_found = False
if self.datetime_column != "index":
datetime_values = X[self.datetime_column]
else:
datetime_values = X.index
if not isinstance(datetime_values, pd.DatetimeIndex):
datetime_values = y.index
if not isinstance(datetime_values, pd.DatetimeIndex):
no_dt_found = True
try:
inferred_freq = pd.infer_freq(datetime_values)
except TypeError:
no_dt_found = True
if no_dt_found:
messages.append(
DataCheckError(
message=f"Datetime information could not be found in the data, or was not in a supported datetime format.",
data_check_name=self.name,
message_code=DataCheckMessageCode.DATETIME_INFORMATION_NOT_FOUND,
).to_dict()
)
return messages
col_name = (
self.datetime_column if self.datetime_column != "index" else "either index"
)
is_increasing = pd.DatetimeIndex(datetime_values).is_monotonic_increasing
nan_columns = datetime_values.isna().any()
if nan_columns:
messages.append(
DataCheckError(
message=f"Input datetime column '{col_name}' contains NaN values. Please impute NaN values or drop these rows.",
data_check_name=self.name,
message_code=DataCheckMessageCode.DATETIME_HAS_NAN,
).to_dict()
)
# NaN values are a more significant error than missing values and will break the data check logic, so we return here
return messages
if not inferred_freq:
# Check for only one row per datetime
duplicate_dates = pd.Series(datetime_values).diff(1) == pd.Timedelta(0)
if duplicate_dates.any():
messages.append(
DataCheckError(
message=f"Column '{col_name}' has more than one row with the same datetime value.",
data_check_name=self.name,
message_code=DataCheckMessageCode.DATETIME_HAS_REDUNDANT_ROW,
).to_dict()
)
# drop any duplicates before continuing
datetime_values = datetime_values[~duplicate_dates]
interval_size = 3
frequencies = []
for i in range(len(datetime_values) - (interval_size - 1)):
freq = pd.infer_freq(datetime_values[i : i + interval_size])
frequencies.append(freq)
num_disruptions = len(set(frequencies))
# Check for no date missing in ordered dates
if num_disruptions > 1 and is_increasing:
messages.append(
DataCheckError(
message=f"Column '{col_name}' has datetime values missing between start and end date.",
data_check_name=self.name,
message_code=DataCheckMessageCode.DATETIME_IS_MISSING_VALUES,
).to_dict()
)
# Give a generic uneven interval error if two or more frequencies are detected
if num_disruptions > 2 or (num_disruptions > 0 and not is_increasing):
messages.append(
DataCheckError(
message=f"No frequency could be detected in column '{col_name}', possibly due to uneven intervals.",
data_check_name=self.name,
message_code=DataCheckMessageCode.DATETIME_HAS_UNEVEN_INTERVALS,
).to_dict()
)
if not is_increasing:
messages.append(
DataCheckError(
message="Datetime values must be sorted in ascending order.",
data_check_name=self.name,
message_code=DataCheckMessageCode.DATETIME_IS_NOT_MONOTONIC,
).to_dict()
)
return messages