Update split_data to call split_multiseries_data

docs/source/release_notes.rst

@@ -10,6 +10,7 @@ Release Notes
  * Extended TimeSeriesRegularizer to support multiseries :pr:`4303`
  * Fixes
  * Changes
+ * Updated ``split_data`` to call ``split_multiseries_data`` when passed stacked multiseries data :pr:`4312`
  * Documentation Changes
  * Removed LightGBM's excessive amount of warnings :pr:`4308`
  * Testing Changes

evalml/preprocessing/utils.py

@@ -4,7 +4,12 @@
 
 from evalml.pipelines.utils import stack_data, stack_X, unstack_multiseries
 from evalml.preprocessing.data_splitters import TrainingValidationSplit
-from evalml.problem_types import is_classification, is_regression, is_time_series
+from evalml.problem_types import (
+ is_classification,
+ is_multiseries,
+ is_regression,
+ is_time_series,
+)
 from evalml.utils import infer_feature_types
 
 
@@ -118,6 +123,9 @@
  Returns:
  pd.DataFrame, pd.DataFrame, pd.Series, pd.Series: Feature and target data each split into train and test sets.
 
+ Raises:
+ ValueError: If the problem_configuration is missing or does not contain both a time_index and series_id for multiseries problems.
+
  Examples:
  >>> X = pd.DataFrame([1, 2, 3, 4, 5, 6], columns=["First"])
  >>> y = pd.Series([8, 9, 10, 11, 12, 13])
@@ -144,6 +152,27 @@
  1 9
  dtype: int64
  """
+ if is_multiseries(problem_type) and isinstance(y, pd.Series):
+ if problem_configuration is None:
+ raise ValueError(
+ "split_data requires problem_configuration for multiseries problems",
+ )
+ series_id = problem_configuration.get("series_id")
+ time_index = problem_configuration.get("time_index")
+ if series_id is None or time_index is None:
+ raise ValueError(
+ "split_data needs both series_id and time_index values in the problem_configuration to split multiseries data",
+ )
+ return split_multiseries_data(
+ X,
+ y,
+ series_id,
+ time_index,
+ problem_configuration=problem_configuration,
+ test_size=test_size,
+ random_seed=random_seed,
+ )
+
  X = infer_feature_types(X)
  y = infer_feature_types(y)
 

evalml/tests/preprocessing_tests/test_split_data.py

@@ -6,6 +6,7 @@
  ProblemTypes,
  is_binary,
  is_multiclass,
+ is_multiseries,
  is_regression,
  is_time_series,
 )
@@ -19,6 +20,7 @@
  X_y_binary,
  X_y_multi,
  X_y_regression,
+ multiseries_ts_data_unstacked,
  make_data_type,
 ):
  if is_binary(problem_type):
@@ -30,6 +32,8 @@
  problem_configuration = None
  if is_time_series(problem_type):
  problem_configuration = {"gap": 1, "max_delay": 7, "time_index": "date"}
+ if is_multiseries(problem_type):
+ X, y = multiseries_ts_data_unstacked
 
  X = make_data_type(data_type, X)
  y = make_data_type(data_type, y)
@@ -50,17 +54,28 @@
  assert len(y_test) == test_size
  assert isinstance(X_train, pd.DataFrame)
  assert isinstance(X_test, pd.DataFrame)
- assert isinstance(y_train, pd.Series)
- assert isinstance(y_test, pd.Series)
+ if not is_multiseries(problem_type):
+ assert isinstance(y_train, pd.Series)
+ assert isinstance(y_test, pd.Series)
+ else:
+ assert isinstance(y_train, pd.DataFrame)
+ assert isinstance(y_test, pd.DataFrame)
+ pd.testing.assert_frame_equal(X_test, X[int(train_size) :], check_dtype=False)
+ pd.testing.assert_frame_equal(y_test, y[int(train_size) :], check_dtype=False)
 
- if is_time_series(problem_type):
+ if is_time_series(problem_type) and not is_multiseries(problem_type):
  pd.testing.assert_frame_equal(X_test, X[int(train_size) :], check_dtype=False)
  pd.testing.assert_series_equal(y_test, y[int(train_size) :], check_dtype=False)
 
 
 @pytest.mark.parametrize("problem_type", ProblemTypes.all_problem_types)
 @pytest.mark.parametrize("data_type", ["np", "pd", "ww"])
-def test_split_data_defaults(problem_type, data_type, get_test_data_from_configuration):
+def test_split_data_defaults(
+ problem_type,
+ data_type,
+ get_test_data_from_configuration,
+ multiseries_ts_data_unstacked,
+):
  X, y = get_test_data_from_configuration(
  data_type,
  problem_type,
@@ -71,6 +86,8 @@
  problem_configuration = None
  if is_time_series(problem_type):
  problem_configuration = {"gap": 1, "max_delay": 7, "time_index": "date"}
+ if is_multiseries(problem_type):
+ X, y = multiseries_ts_data_unstacked
  test_pct = 0.1
  else:
  test_pct = 0.2
@@ -92,7 +109,18 @@
  X = pd.DataFrame(X)
  y = pd.Series(y)
  pd.testing.assert_frame_equal(X_test, X[int(train_size) :], check_dtype=False)
- pd.testing.assert_series_equal(y_test, y[int(train_size) :], check_dtype=False)
+ if not is_multiseries(problem_type):
+ pd.testing.assert_series_equal(
+ y_test,
+ y[int(train_size) :],
+ check_dtype=False,
+ )
+ else:
+ pd.testing.assert_frame_equal(
+ y_test,
+ y[int(train_size) :],
+ check_dtype=False,
+ )
 
 
 @pytest.mark.parametrize("test", ["fh_limitation", "no_fh_limitation"])
@@ -127,8 +155,33 @@
  assert len(y_test) == test_size
 
 
+def test_split_data_calls_multiseries_error(multiseries_ts_data_stacked):
+ X, y = multiseries_ts_data_stacked
+ with pytest.raises(
+ ValueError,
+ match="requires problem_configuration for multiseries",
+ ):
+ split_data(X, y, problem_type="multiseries time series regression")
+
+ with pytest.raises(
+ ValueError,
+ match="needs both series_id and time_index values in the problem_configuration",
+ ):
+ split_data(
+ X,
+ y,
+ problem_type="multiseries time series regression",
+ problem_configuration={"time_index": "date"},
+ )
+
+
 @pytest.mark.parametrize("no_features", [True, False])
-def test_split_multiseries_data(no_features, multiseries_ts_data_stacked):
+@pytest.mark.parametrize("splitting_function", ["split_data", "split_multiseries_data"])
+def test_split_multiseries_data(
+ no_features,
+ splitting_function,
+ multiseries_ts_data_stacked,
+):
  X, y = multiseries_ts_data_stacked
 
  if no_features:
@@ -137,12 +190,22 @@
  X_train_expected, X_holdout_expected = X[:-10], X[-10:]
  y_train_expected, y_holdout_expected = y[:-10], y[-10:]
 
- X_train, X_holdout, y_train, y_holdout = split_multiseries_data(
- X,
- y,
- "series_id",
- "date",
- )
+ # Results should be identical whether split_multiseries_data is called through
+ # split_data or directly
+ if splitting_function == "split_data":
+ X_train, X_holdout, y_train, y_holdout = split_data(
+ X,
+ y,
+ problem_type="multiseries time series regression",
+ problem_configuration={"time_index": "date", "series_id": "series_id"},
+ )
+ else:
+ X_train, X_holdout, y_train, y_holdout = split_multiseries_data(
+ X,
+ y,
+ "series_id",
+ "date",
+ )
 
  pd.testing.assert_frame_equal(
  X_train.sort_index(axis=1),