TargetAgg¶
Computes a streaming aggregate of the target values.
This transformer is identical to feature_extraction.Agg
, the only difference is that it operates on the target rather than on a feature. At each step, the running statistic how
of target values in group by
is updated with the target. It is therefore a supervised transformer.
Parameters¶
-
by
Type → str | list[str] | None
The feature by which to group the target values. All the data is included in the aggregate if this is
None
. -
how
Type → stats.base.Univariate | utils.Rolling | utils.TimeRolling
The statistic to compute.
-
target_name
Default →
y
The target name which is used in the result.
Attributes¶
-
state
Return the current values for each group as a series.
-
target_name
Examples¶
Consider the following dataset, where the second value of each value is the target:
dataset = [
({'country': 'France', 'place': 'Taco Bell'}, 42),
({'country': 'Sweden', 'place': 'Burger King'}, 16),
({'country': 'France', 'place': 'Burger King'}, 24),
({'country': 'Sweden', 'place': 'Taco Bell'}, 58),
({'country': 'Sweden', 'place': 'Burger King'}, 20),
({'country': 'France', 'place': 'Taco Bell'}, 50),
({'country': 'France', 'place': 'Burger King'}, 10),
({'country': 'Sweden', 'place': 'Taco Bell'}, 80)
]
As an example, let's perform a target encoding of the place
feature. Instead of simply
updating a running average, we use a stats.BayesianMean
which allows us to incorporate
some prior knowledge. This makes subsequent models less prone to overfitting. Indeed, it
dampens the fact that too few samples might have been seen within a group.
from river import feature_extraction
from river import stats
agg = feature_extraction.TargetAgg(
by='place',
how=stats.BayesianMean(
prior=3,
prior_weight=1
)
)
for x, y in dataset:
print(agg.transform_one(x))
agg.learn_one(x, y)
{'y_bayes_mean_by_place': 3.0}
{'y_bayes_mean_by_place': 3.0}
{'y_bayes_mean_by_place': 9.5}
{'y_bayes_mean_by_place': 22.5}
{'y_bayes_mean_by_place': 14.333}
{'y_bayes_mean_by_place': 34.333}
{'y_bayes_mean_by_place': 15.75}
{'y_bayes_mean_by_place': 38.25}
Just like with feature_extraction.Agg
, we can specify multiple features on which to
group the data:
agg = feature_extraction.TargetAgg(
by=['place', 'country'],
how=stats.BayesianMean(
prior=3,
prior_weight=1
)
)
for x, y in dataset:
print(agg.transform_one(x))
agg.learn_one(x, y)
{'y_bayes_mean_by_place_and_country': 3.0}
{'y_bayes_mean_by_place_and_country': 3.0}
{'y_bayes_mean_by_place_and_country': 3.0}
{'y_bayes_mean_by_place_and_country': 3.0}
{'y_bayes_mean_by_place_and_country': 9.5}
{'y_bayes_mean_by_place_and_country': 22.5}
{'y_bayes_mean_by_place_and_country': 13.5}
{'y_bayes_mean_by_place_and_country': 30.5}
agg.state
place country
Taco Bell France 31.666667
Burger King Sweden 13.000000
France 12.333333
Taco Bell Sweden 47.000000
Name: y_bayes_mean_by_place_and_country, dtype: float64
This transformer can also be used in conjunction with utils.TimeRolling
. The latter requires
a t
argument, which is a timestamp that indicates when the current row was observed. For
instance, we can calculate the average (how) revenue (on) for each place (by) over the last
7 days (t):
import datetime as dt
import random
import string
from river import utils
agg = feature_extraction.TargetAgg(
by="group",
how=utils.TimeRolling(stats.Mean(), dt.timedelta(days=7))
)
for day in range(366):
g = random.choice(string.ascii_lowercase)
x = {"group": g}
y = string.ascii_lowercase.index(g) + random.random()
t = dt.datetime(2023, 1, 1) + dt.timedelta(days=day)
agg.learn_one(x, y, t=t)
Methods¶
learn_one
Update with a set of features x
and a target y
.
Parameters
- x — 'dict'
- y — 'base.typing.Target'
- t — defaults to
None
transform_one
Transform a set of features x
.
Parameters
- x — 'dict'
Returns
dict: The transformed values.