import numpy as np
import pandas as pd
from woodwork import init_series
from woodwork.column_schema import ColumnSchema
from woodwork.logical_types import Datetime, Double
from featuretools.primitives.base.transform_primitive_base import TransformPrimitive
from featuretools.primitives.standard.transform.time_series.utils import (
apply_rolling_agg_to_series,
)
[docs]class RollingOutlierCount(TransformPrimitive):
"""Determines how many values are outliers over a given window.
Description:
Given a list of numbers and a corresponding list of
datetimes, return a rolling count of outliers within the numeric values,
starting at the row `gap` rows away from the current row and looking backward
over the specified window (by `window_length` and `gap`). Values are deemed
outliers using the IQR method, computed over the whole series.
Input datetimes should be monotonic.
Args:
window_length (int, string, optional): Specifies the amount of data included in each window.
If an integer is provided, it will correspond to a number of rows. For data with a uniform sampling
frequency, for example of one day, the window_length will correspond to a period of time, in this case,
7 days for a window_length of 7.
If a string is provided, it must be one of Pandas' offset alias strings ('1D', '1H', etc),
and it will indicate a length of time that each window should span.
The list of available offset aliases can be found at
https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases.
Defaults to 3.
gap (int, string, optional): Specifies a gap backwards from each instance before the
window of usable data begins. If an integer is provided, it will correspond to a number of rows.
If a string is provided, it must be one of Pandas' offset alias strings ('1D', '1H', etc),
and it will indicate a length of time between a target instance and the beginning of its window.
Defaults to 1, which excludes the target instance from the window.
min_periods (int, optional): Minimum number of observations required for performing calculations
over the window. Can only be as large as window_length when window_length is an integer.
When window_length is an offset alias string, this limitation does not exist, but care should be taken
to not choose a min_periods that will always be larger than the number of observations in a window.
Defaults to 1.
Note:
Only offset aliases with fixed frequencies can be used when defining gap and window_length.
This means that aliases such as `M` or `W` cannot be used, as they can indicate different
numbers of days. ('M', because different months are different numbers of days;
'W' because week will indicate a certain day of the week, like W-Wed, so that will
indicate a different number of days depending on the anchoring date.)
Note:
When using an offset alias to define `gap`, an offset alias must also be used to define `window_length`.
This limitation does not exist when using an offset alias to define `window_length`. In fact,
if the data has a uniform sampling frequency, it is preferable to use a numeric `gap` as it is more
efficient.
Examples:
>>> import pandas as pd
>>> rolling_outlier_count = RollingOutlierCount(window_length=4)
>>> times = pd.date_range(start='2019-01-01', freq='1min', periods=6)
>>> rolling_outlier_count(times, [0, 0, 0, 0, 10, 0]).tolist()
[nan, 0.0, 0.0, 0.0, 0.0, 1.0]
We can also control the gap before the rolling calculation.
>>> import pandas as pd
>>> rolling_outlier_count = RollingOutlierCount(window_length=4, gap=0)
>>> times = pd.date_range(start='2019-01-01', freq='1min', periods=6)
>>> rolling_outlier_count(times, [0, 0, 0, 0, 10, 0]).tolist()
[0.0, 0.0, 0.0, 0.0, 1.0, 1.0]
We can also control the minimum number of periods required for the rolling calculation.
>>> import pandas as pd
>>> rolling_outlier_count = RollingOutlierCount(window_length=4, min_periods=3)
>>> times = pd.date_range(start='2019-01-01', freq='1min', periods=6)
>>> rolling_outlier_count(times, [0, 0, 0, 0, 10, 0]).tolist()
[nan, nan, nan, 0.0, 0.0, 1.0]
We can also set the window_length and gap using offset alias strings.
>>> import pandas as pd
>>> rolling_outlier_count = RollingOutlierCount(window_length='4min', gap='1min')
>>> times = pd.date_range(start='2019-01-01', freq='1min', periods=6)
>>> rolling_outlier_count(times, [0, 0, 0, 0, 10, 0]).tolist()
[nan, 0.0, 0.0, 0.0, 0.0, 1.0]
"""
name = "rolling_outlier_count"
input_types = [
ColumnSchema(logical_type=Datetime, semantic_tags={"time_index"}),
ColumnSchema(semantic_tags={"numeric"}),
]
return_type = ColumnSchema(logical_type=Double, semantic_tags={"numeric"})
uses_full_dataframe = True
[docs] def __init__(self, window_length=3, gap=1, min_periods=0):
self.window_length = window_length
self.gap = gap
self.min_periods = min_periods
def get_outliers_count(self, numeric_series):
# We know the column is numeric, so use the Double logical type in case Woodwork's
# type inference could not infer a numeric type
if not len(numeric_series.dropna()):
return np.nan
if numeric_series.ww.schema is None:
numeric_series = init_series(numeric_series, logical_type="Double")
box_plot_info = numeric_series.ww.box_plot_dict()
return len(box_plot_info["high_values"]) + len(box_plot_info["low_values"])
def get_function(self):
def rolling_outlier_count(datetime, numeric):
x = pd.Series(numeric.values, index=datetime.values)
return apply_rolling_agg_to_series(
series=x,
agg_func=self.get_outliers_count,
window_length=self.window_length,
gap=self.gap,
min_periods=self.min_periods,
ignore_window_nans=False,
)
return rolling_outlier_count