"""Pipeline base class for time-series problems."""
import pandas as pd
import woodwork as ww
from evalml.pipelines import PipelineBase
from evalml.pipelines.pipeline_meta import PipelineBaseMeta
from evalml.utils import infer_feature_types
from evalml.utils.gen_utils import are_datasets_separated_by_gap_time_index
[docs]class TimeSeriesPipelineBase(PipelineBase, metaclass=PipelineBaseMeta):
"""Pipeline base class for time series problems.
Args:
component_graph (ComponentGraph, list, dict): ComponentGraph instance, list of components in order, or dictionary of components.
Accepts strings or ComponentBase subclasses in the list.
Note that when duplicate components are specified in a list, the duplicate component names will be modified with the
component's index in the list. For example, the component graph
[Imputer, One Hot Encoder, Imputer, Logistic Regression Classifier] will have names
["Imputer", "One Hot Encoder", "Imputer_2", "Logistic Regression Classifier"]
parameters (dict): Dictionary with component names as keys and dictionary of that component's parameters as values.
An empty dictionary {} implies using all default values for component parameters. Pipeline-level
parameters such as time_index, gap, and max_delay must be specified with the "pipeline" key. For example:
Pipeline(parameters={"pipeline": {"time_index": "Date", "max_delay": 4, "gap": 2}}).
random_seed (int): Seed for the random number generator. Defaults to 0.
"""
def __init__(
self,
component_graph,
parameters=None,
custom_name=None,
random_seed=0,
):
if not parameters or "pipeline" not in parameters:
raise ValueError(
"time_index, gap, max_delay, and forecast_horizon parameters cannot be omitted from the parameters dict. "
"Please specify them as a dictionary with the key 'pipeline'.",
)
self.pipeline_params = parameters["pipeline"]
self.gap = self.pipeline_params["gap"]
self.max_delay = self.pipeline_params["max_delay"]
self.forecast_horizon = self.pipeline_params["forecast_horizon"]
self.time_index = self.pipeline_params["time_index"]
self.frequency = None
if self.time_index is None:
raise ValueError("Parameter time_index cannot be None!")
super().__init__(
component_graph,
custom_name=custom_name,
parameters=parameters,
random_seed=random_seed,
)
datetime_featurizer_included = (
"DateTime Featurizer" in self.component_graph.compute_order
or "Not Known In Advance Pipeline - DateTime Featurizer"
in self.component_graph.compute_order
)
time_series_featurizer_included = (
"Time Series Featurizer" in self.component_graph.compute_order
or "Not Known In Advance Pipeline - Time Series Featurizer"
in self.component_graph.compute_order
)
time_series_native_estimators = [
"ARIMA Regressor",
"Prophet Regressor",
"VARMAX Regressor",
]
self.should_skip_featurization = (
not datetime_featurizer_included
and not time_series_featurizer_included
and self.estimator is not None
)
self.should_drop_time_index = (
self.should_skip_featurization
and self.estimator.name not in time_series_native_estimators
)
@staticmethod
def _convert_to_woodwork(X, y):
if X is None:
X = pd.DataFrame()
X = infer_feature_types(X)
y = infer_feature_types(y)
return X, y
@staticmethod
def _move_index_forward(index, gap):
"""Fill in the index of the gap features and values with the right values."""
if isinstance(index, (pd.DatetimeIndex, pd.PeriodIndex, pd.TimedeltaIndex)):
return index.shift(gap)
else:
return index + gap
def _add_training_data_to_X_Y(self, X, y, X_train, y_train):
"""Append the training data to the holdout data.
Need to do this so that we have all the data we need to compute lagged features on the holdout set.
"""
last_row_of_training = self.forecast_horizon + self.max_delay + self.gap
gap_features = pd.DataFrame()
gap_target = pd.Series() if isinstance(y, pd.Series) else pd.DataFrame()
if (
are_datasets_separated_by_gap_time_index(
X_train,
X,
self.pipeline_params,
self.frequency,
)
and self.gap
):
# The training data does not have the gap dates so don't need to include them
last_row_of_training -= self.gap
# Instead, we'll create some dummy data to represent the missing gap dates
# These do not show up in the features used for prediction
gap_features = X_train.iloc[[-1] * self.gap]
gap_features.index = self._move_index_forward(
X_train.index[-self.gap :],
self.gap,
)
gap_target = y_train.iloc[[-1] * self.gap]
gap_target.index = self._move_index_forward(
y_train.index[-self.gap :],
self.gap,
)
# Properly fill in the dates in the gap
time_index = self.pipeline_params["time_index"]
correct_range = pd.date_range(
start=X_train[time_index].iloc[-1],
periods=self.gap + 1,
freq=self.frequency,
)[1:]
gap_features[time_index] = correct_range
features_to_concat = [
X_train.iloc[-last_row_of_training:],
gap_features,
X,
]
targets_to_concat = [
y_train.iloc[-last_row_of_training:],
gap_target,
y,
]
padded_features = pd.concat(features_to_concat, axis=0).fillna(method="ffill")
padded_target = pd.concat(targets_to_concat, axis=0).fillna(method="ffill")
padded_features.ww.init(schema=X_train.ww.schema)
if isinstance(padded_target, pd.Series):
padded_target = ww.init_series(
padded_target,
logical_type=y_train.ww.logical_type,
)
else: # Multiseries case
padded_target.ww.init(schema=y_train.ww.schema)
return padded_features, padded_target
def _drop_time_index(self, X, y):
"""Helper method to drop the time index column from the data if DateTime Featurizer is not present."""
if self.should_drop_time_index and self.time_index in X.columns:
index_name = X.index.name
time_index = pd.DatetimeIndex(X[self.time_index], freq=self.frequency)
y_schema = y.ww.schema
y = y.set_axis(time_index)
y.ww.init(schema=y_schema)
if X.ww.schema is not None:
X = X.ww.copy()
X.ww.set_time_index(None)
X.ww.set_index(self.time_index)
X = X.ww.drop(self.time_index)
X.index.freq = time_index.freq
else:
X.set_index(time_index)
X = X.drop(self.time_index, axis=1)
X.index.name = index_name
y.index.name = index_name
return X, y
[docs] def predict_in_sample(
self,
X,
y,
X_train,
y_train,
objective=None,
calculating_residuals=False,
):
"""Predict on future data where the target is known, e.g. cross validation.
Args:
X (pd.DataFrame or np.ndarray): Future data of shape [n_samples, n_features]
y (pd.Series, np.ndarray): Future target of shape [n_samples]
X_train (pd.DataFrame, np.ndarray): Data the pipeline was trained on of shape [n_samples_train, n_feautures]
y_train (pd.Series, np.ndarray): Targets used to train the pipeline of shape [n_samples_train]
objective (ObjectiveBase, str, None): Objective used to threshold predicted probabilities, optional.
calculating_residuals (bool): Whether we're calling predict_in_sample to calculate the residuals. This means
the X and y arguments are not future data, but actually the train data.
Returns:
pd.Series: Estimated labels.
Raises:
ValueError: If final component is not an Estimator.
"""
if self.estimator is None:
raise ValueError(
"Cannot call predict_in_sample() on a component graph because the final component is not an Estimator.",
)
X, y = self._drop_time_index(X, y)
X_train, y_train = self._drop_time_index(X_train, y_train)
X, y = self._ensure_correct_indices(X, y, X_train)
target = infer_feature_types(y)
features = self.transform_all_but_final(
X,
target,
X_train,
y_train,
calculating_residuals=calculating_residuals,
)
predictions = self._estimator_predict(features)
if len(predictions) == len(y):
predictions.index = y.index
predictions = self.inverse_transform(predictions)
if isinstance(predictions, pd.Series):
predictions = predictions.rename(self.input_target_name)
return infer_feature_types(predictions)
def _ensure_correct_indices(self, X, y, X_train):
"""Ensures that X and y holdout's indices are the correct integer or time units w.r.t the training data.
For predict in sample where the holdout is known to follow the training data.
"""
if X_train.index.is_numeric():
starting_index = X_train.index[-1] + 1 + self.gap
correct_index = range(starting_index, starting_index + len(y))
X.index = correct_index
y.index = correct_index
return X, y
def _create_empty_series(self, y_train, size):
return ww.init_series(
pd.Series([y_train.iloc[0]] * size, name=y_train.name),
logical_type=y_train.ww.logical_type,
)
[docs] def predict(self, X, objective=None, X_train=None, y_train=None):
"""Predict on future data where target is not known.
Args:
X (pd.DataFrame, or np.ndarray): Data of shape [n_samples, n_features].
objective (Object or string): The objective to use to make predictions.
X_train (pd.DataFrame or np.ndarray or None): Training data.
y_train (pd.Series or None): Training labels.
Raises:
ValueError: If X_train and/or y_train are None or if final component is not an Estimator.
Returns:
Predictions.
"""
if X_train is None:
raise ValueError(
"Make sure to include an input for X_train when calling time series' predict",
)
elif y_train is None:
raise ValueError(
"Make sure to include an input for y_train when calling time series' predict",
)
if self.estimator is None:
raise ValueError(
"Cannot call predict() on a component graph because the final component is not an Estimator.",
)
X = infer_feature_types(X)
X.index = self._move_index_forward(
X_train.index[-X.shape[0] :],
self.gap + X.shape[0],
)
X, y = self._drop_time_index(X, pd.Series([0] * len(X)))
X_train, y_train = self._drop_time_index(X_train, y_train)
X_train, y_train = self._convert_to_woodwork(X_train, y_train)
y_holdout = self._create_empty_series(y_train, X.shape[0])
y_holdout = infer_feature_types(y_holdout)
y_holdout.index = X.index
return self.predict_in_sample(
X,
y_holdout,
X_train,
y_train,
objective=objective,
)
def _estimator_predict(self, features):
"""Get estimator predictions.
This helper passes y as an argument if needed by the estimator.
"""
return self.estimator.predict(features)
[docs] def dates_needed_for_prediction(self, date):
"""Return dates needed to forecast the given date in the future.
Args:
date (pd.Timestamp): Date to forecast in the future.
Returns:
dates_needed (tuple(pd.Timestamp)): Range of dates needed to forecast the given date.
"""
beginning_date_num = (
self.forecast_horizon
+ self.max_delay # include start delay for featurization
+ self.gap # add first gap for the actual gap from the end date
+ self.gap # add another gap to ensure training data is greater than gap
+ 1 # for the + 1 in the time series featurizer
)
beginning_date = date - pd.tseries.frequencies.to_offset(
f"{beginning_date_num}{self.frequency}",
)
end_date_num = 1 + self.gap
end_date = date - pd.tseries.frequencies.to_offset(
f"{end_date_num}{self.frequency}",
)
return (beginning_date, end_date)
[docs] def dates_needed_for_prediction_range(self, start_date, end_date):
"""Return dates needed to forecast the given date in the future.
Args:
start_date (pd.Timestamp): Start date of range to forecast in the future.
end_date (pd.Timestamp): End date of range to forecast in the future.
Returns:
dates_needed (tuple(pd.Timestamp)): Range of dates needed to forecast the given date.
Raises:
ValueError: If start_date doesn't come before end_date
"""
if start_date > end_date:
raise ValueError("`start_date` must come before `end_date`.")
start_range = self.dates_needed_for_prediction(start_date)
end_range = self.dates_needed_for_prediction(end_date)
return (start_range[0], end_range[1])