autogluon.timeseries 1.2.1b20250224__py3-none-any.whl → 1.4.1b20251215__py3-none-any.whl

autogluon/timeseries/models/ensemble/ensemble_selection.py

@@ -0,0 +1,167 @@
+import copy
+
+import numpy as np
+
+import autogluon.core as ag
+from autogluon.core.models.greedy_ensemble.ensemble_selection import EnsembleSelection
+from autogluon.timeseries import TimeSeriesDataFrame
+from autogluon.timeseries.metrics import TimeSeriesScorer
+from autogluon.timeseries.utils.datetime import get_seasonality
+
+
+class TimeSeriesEnsembleSelection(EnsembleSelection):
+    def __init__(
+        self,
+        ensemble_size: int,
+        metric: TimeSeriesScorer,
+        problem_type: str = ag.constants.QUANTILE,
+        sorted_initialization: bool = False,
+        bagging: bool = False,
+        tie_breaker: str = "random",
+        random_state: np.random.RandomState | None = None,
+        prediction_length: int = 1,
+        target: str = "target",
+        **kwargs,
+    ):
+        super().__init__(
+            ensemble_size=ensemble_size,
+            metric=metric,  # type: ignore
+            problem_type=problem_type,
+            sorted_initialization=sorted_initialization,
+            bagging=bagging,
+            tie_breaker=tie_breaker,
+            random_state=random_state,
+            **kwargs,
+        )
+        self.prediction_length = prediction_length
+        self.target = target
+        self.metric: TimeSeriesScorer
+
+        self.dummy_pred_per_window = []
+        self.scorer_per_window = []
+
+        self.dummy_pred_per_window: list[TimeSeriesDataFrame] | None
+        self.scorer_per_window: list[TimeSeriesScorer] | None
+        self.data_future_per_window: list[TimeSeriesDataFrame] | None
+
+    def fit(  # type: ignore
+        self,
+        predictions: list[list[TimeSeriesDataFrame]],
+        labels: list[TimeSeriesDataFrame],
+        time_limit: float | None = None,
+    ):
+        return super().fit(
+            predictions=predictions,  # type: ignore
+            labels=labels,  # type: ignore
+            time_limit=time_limit,
+        )
+
+    def _fit(  # type: ignore
+        self,
+        predictions: list[list[TimeSeriesDataFrame]],
+        labels: list[TimeSeriesDataFrame],
+        time_limit: float | None = None,
+        sample_weight: list[float] | None = None,
+    ):
+        # Stack predictions for each model into a 3d tensor of shape [num_val_windows, num_rows, num_cols]
+        stacked_predictions = [np.stack(preds) for preds in predictions]
+
+        self.dummy_pred_per_window = []
+        self.scorer_per_window = []
+        self.data_future_per_window = []
+
+        seasonal_period = self.metric.seasonal_period
+        if seasonal_period is None:
+            seasonal_period = get_seasonality(labels[0].freq)
+
+        for window_idx, data in enumerate(labels):
+            dummy_pred = copy.deepcopy(predictions[0][window_idx])
+            # This should never happen; sanity check to make sure that all predictions have the same index
+            assert all(dummy_pred.index.equals(pred[window_idx].index) for pred in predictions)
+            assert all(dummy_pred.columns.equals(pred[window_idx].columns) for pred in predictions)
+
+            self.dummy_pred_per_window.append(dummy_pred)
+
+            scorer = copy.deepcopy(self.metric)
+            # Split the observed time series once to avoid repeated computations inside the evaluator
+            data_past = data.slice_by_timestep(None, -self.prediction_length)
+            data_future = data.slice_by_timestep(-self.prediction_length, None)
+            scorer.save_past_metrics(data_past, target=self.target, seasonal_period=seasonal_period)
+            self.scorer_per_window.append(scorer)
+            self.data_future_per_window.append(data_future)
+
+        super()._fit(
+            predictions=stacked_predictions,
+            labels=data_future,  # type: ignore
+            time_limit=time_limit,
+        )
+        self.dummy_pred_per_window = None
+        self.evaluator_per_window = None
+        self.data_future_per_window = None
+
+    def _calculate_regret(  # type: ignore
+        self,
+        y_true,
+        y_pred_proba,
+        metric: TimeSeriesScorer,
+        sample_weight=None,
+    ):
+        # Compute average score across all validation windows
+        total_score = 0.0
+
+        assert self.data_future_per_window is not None
+        assert self.dummy_pred_per_window is not None
+        assert self.scorer_per_window is not None
+
+        for window_idx, data_future in enumerate(self.data_future_per_window):
+            dummy_pred = self.dummy_pred_per_window[window_idx]
+            dummy_pred[list(dummy_pred.columns)] = y_pred_proba[window_idx]
+            # We use scorer.compute_metric instead of scorer.score to avoid repeated calls to scorer.save_past_metrics
+            metric_value = self.scorer_per_window[window_idx].compute_metric(
+                data_future,
+                dummy_pred,
+                target=self.target,
+            )
+            total_score += metric.sign * metric_value
+        avg_score = total_score / len(self.data_future_per_window)
+        # score: higher is better, regret: lower is better, so we flip the sign
+        return -avg_score
+
+
+def fit_time_series_ensemble_selection(
+    data_per_window: list[TimeSeriesDataFrame],
+    predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+    ensemble_size: int,
+    eval_metric: TimeSeriesScorer,
+    prediction_length: int = 1,
+    target: str = "target",
+    time_limit: float | None = None,
+) -> dict[str, float]:
+    """Fit ensemble selection for time series forecasting and return ensemble weights.
+
+    Parameters
+    ----------
+    data_per_window:
+        List of ground truth time series data for each validation window.
+    predictions_per_window:
+        Dictionary mapping model names to their predictions for each validation window.
+    ensemble_size:
+        Number of iterations of the ensemble selection algorithm.
+
+    Returns
+    -------
+    weights:
+        Dictionary mapping the model name to its weight in the ensemble.
+    """
+    ensemble_selection = TimeSeriesEnsembleSelection(
+        ensemble_size=ensemble_size,
+        metric=eval_metric,
+        prediction_length=prediction_length,
+        target=target,
+    )
+    ensemble_selection.fit(
+        predictions=list(predictions_per_window.values()),
+        labels=data_per_window,
+        time_limit=time_limit,
+    )
+    return {model: float(weight) for model, weight in zip(predictions_per_window.keys(), ensemble_selection.weights_)}

autogluon/timeseries/models/ensemble/per_item_greedy.py

@@ -0,0 +1,172 @@
+import logging
+import pprint
+import time
+from typing import Any
+
+import pandas as pd
+from joblib import Parallel, delayed
+
+from autogluon.timeseries import TimeSeriesDataFrame
+from autogluon.timeseries.utils.constants import AG_DEFAULT_N_JOBS
+
+from .abstract import AbstractTimeSeriesEnsembleModel
+from .ensemble_selection import fit_time_series_ensemble_selection
+
+logger = logging.getLogger(__name__)
+
+
+class PerItemGreedyEnsemble(AbstractTimeSeriesEnsembleModel):
+    """Per-item greedy ensemble that fits separate weighted ensembles for each individual time series.
+
+    This ensemble applies the greedy Ensemble Selection algorithm by Caruana et al. [Car2004]_ independently
+    to each time series in the dataset, allowing for customized model combinations that adapt to the
+    specific characteristics of individual series. Each time series gets its own optimal ensemble weights
+    based on predictions for that particular series. If items not seen during training are provided at prediction
+    time, average model weight across the training items will be used for their predictions.
+
+    The per-item approach is particularly effective for datasets with heterogeneous time series that
+    exhibit different patterns, seasonalities, or noise characteristics.
+
+    The algorithm uses parallel processing to efficiently fit ensembles across all time series.
+
+    Other Parameters
+    ----------------
+    ensemble_size : int, default = 100
+        Number of models (with replacement) to include in the ensemble.
+    n_jobs : int or float, default = joblib.cpu_count(only_physical_cores=True)
+        Number of CPU cores used to fit the ensembles in parallel.
+
+    References
+    ----------
+    .. [Car2004] Caruana, Rich, et al. "Ensemble selection from libraries of models."
+        Proceedings of the twenty-first international conference on Machine learning. 2004.
+    """
+
+    def __init__(self, name: str | None = None, **kwargs):
+        if name is None:
+            name = "PerItemWeightedEnsemble"
+        super().__init__(name=name, **kwargs)
+        self.weights_df: pd.DataFrame
+        self.average_weight: pd.Series
+
+    @property
+    def model_names(self) -> list[str]:
+        return list(self.weights_df.columns)
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        return {"ensemble_size": 100, "n_jobs": AG_DEFAULT_N_JOBS}
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ) -> None:
+        model_names = list(predictions_per_window.keys())
+        item_ids = data_per_window[0].item_ids
+        n_jobs = min(self.get_hyperparameter("n_jobs"), len(item_ids))
+
+        predictions_per_item = self._split_predictions_per_item(predictions_per_window)
+        data_per_item = self._split_data_per_item(data_per_window)
+
+        ensemble_selection_kwargs = dict(
+            ensemble_size=self.get_hyperparameter("ensemble_size"),
+            eval_metric=self.eval_metric,
+            prediction_length=self.prediction_length,
+            target=self.target,
+        )
+
+        time_limit_per_item = None if time_limit is None else time_limit * n_jobs / len(item_ids)
+        end_time = None if time_limit is None else time.time() + time_limit
+
+        # Fit ensemble for each item in parallel
+        executor = Parallel(n_jobs=n_jobs)
+        weights_per_item = executor(
+            delayed(self._fit_item_ensemble)(
+                data_per_item[item_id],
+                predictions_per_item[item_id],
+                time_limit_per_item=time_limit_per_item,
+                end_time=end_time,
+                **ensemble_selection_kwargs,
+            )
+            for item_id in item_ids
+        )
+        self.weights_df = pd.DataFrame(weights_per_item, index=item_ids, columns=model_names)  # type: ignore
+        self.average_weight = self.weights_df.mean(axis=0)
+
+        # Drop models with zero average weight
+        if (self.average_weight == 0).any():
+            models_to_keep = self.average_weight[self.average_weight > 0].index
+            self.weights_df = self.weights_df[models_to_keep]
+            self.average_weight = self.average_weight[models_to_keep]
+
+        weights_for_printing = {model: round(float(weight), 2) for model, weight in self.average_weight.items()}
+        logger.info(f"\tAverage ensemble weights: {pprint.pformat(weights_for_printing, width=200)}")
+
+    def _split_predictions_per_item(
+        self, predictions_per_window: dict[str, list[TimeSeriesDataFrame]]
+    ) -> dict[str, dict[str, list[TimeSeriesDataFrame]]]:
+        """Build a dictionary mapping item_id -> dict[model_name, list[TimeSeriesDataFrame]]."""
+        item_ids = list(predictions_per_window.values())[0][0].item_ids
+
+        predictions_per_item = {}
+        for i, item_id in enumerate(item_ids):
+            item_predictions = {}
+            for model_name, preds_per_window in predictions_per_window.items():
+                item_preds_per_window = [
+                    pred.iloc[i * self.prediction_length : (i + 1) * self.prediction_length]
+                    for pred in preds_per_window
+                ]
+                item_predictions[model_name] = item_preds_per_window
+            predictions_per_item[item_id] = item_predictions
+        return predictions_per_item
+
+    def _split_data_per_item(self, data_per_window: list[TimeSeriesDataFrame]) -> dict[str, list[TimeSeriesDataFrame]]:
+        """Build a dictionary mapping item_id -> ground truth values across all windows."""
+        item_ids = data_per_window[0].item_ids
+        data_per_item = {item_id: [] for item_id in item_ids}
+
+        for data in data_per_window:
+            indptr = data.get_indptr()
+            for item_idx, item_id in enumerate(item_ids):
+                new_slice = data.iloc[indptr[item_idx] : indptr[item_idx + 1]]
+                data_per_item[item_id].append(new_slice)
+        return data_per_item
+
+    @staticmethod
+    def _fit_item_ensemble(
+        data_per_window: list[TimeSeriesDataFrame],
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        time_limit_per_item: float | None = None,
+        end_time: float | None = None,
+        **ensemble_selection_kwargs,
+    ) -> dict[str, float]:
+        """Fit ensemble for a single item."""
+        if end_time is not None:
+            assert time_limit_per_item is not None
+            time_left = end_time - time.time()
+            time_limit_per_item = min(time_limit_per_item, time_left)
+        return fit_time_series_ensemble_selection(
+            data_per_window, predictions_per_window, time_limit=time_limit_per_item, **ensemble_selection_kwargs
+        )
+
+    def _predict(self, data: dict[str, TimeSeriesDataFrame], **kwargs) -> TimeSeriesDataFrame:
+        assert all(model in data for model in self.weights_df.columns)
+        item_ids = list(data.values())[0].item_ids
+        unseen_item_ids = set(item_ids) - set(self.weights_df.index)
+        if unseen_item_ids:
+            logger.debug(f"Using average weights for {len(unseen_item_ids)} unseen items")
+        weights = self.weights_df.reindex(item_ids).fillna(self.average_weight)
+
+        result = None
+        for model_name in self.weights_df.columns:
+            model_pred = data[model_name]
+            model_weights = weights[model_name].to_numpy().repeat(self.prediction_length)
+            weighted_pred = model_pred.to_data_frame().multiply(model_weights, axis=0)
+            result = weighted_pred if result is None else result + weighted_pred
+
+        return TimeSeriesDataFrame(result)  # type: ignore
+
+    def remap_base_models(self, model_refit_map: dict[str, str]) -> None:
+        self.weights_df.rename(columns=model_refit_map, inplace=True)

autogluon/timeseries/models/ensemble/weighted/__init__.py

@@ -0,0 +1,8 @@
+from .basic import PerformanceWeightedEnsemble, SimpleAverageEnsemble
+from .greedy import GreedyEnsemble
+
+__all__ = [
+    "SimpleAverageEnsemble",
+    "PerformanceWeightedEnsemble",
+    "GreedyEnsemble",
+]

autogluon/timeseries/models/ensemble/weighted/abstract.py

@@ -0,0 +1,45 @@
+import functools
+from abc import ABC
+
+import numpy as np
+
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
+
+from ..abstract import AbstractTimeSeriesEnsembleModel
+
+
+class AbstractWeightedTimeSeriesEnsembleModel(AbstractTimeSeriesEnsembleModel, ABC):
+    """Abstract base class for weighted ensemble models that assign global weights to base models.
+
+    Weighted ensembles combine predictions from multiple base models using learned or computed weights,
+    where each base model receives a single global weight applied across all time series and forecast
+    horizons. The final prediction is computed as a weighted linear combination of base model forecasts.
+    """
+
+    def __init__(self, name: str | None = None, **kwargs):
+        super().__init__(name=name, **kwargs)
+        self.model_to_weight: dict[str, float] = {}
+
+    @property
+    def model_names(self) -> list[str]:
+        return list(self.model_to_weight.keys())
+
+    @property
+    def model_weights(self) -> np.ndarray:
+        return np.array(list(self.model_to_weight.values()), dtype=np.float64)
+
+    def _predict(self, data: dict[str, TimeSeriesDataFrame], **kwargs) -> TimeSeriesDataFrame:
+        weighted_predictions = [data[model_name] * weight for model_name, weight in self.model_to_weight.items()]
+        return functools.reduce(lambda x, y: x + y, weighted_predictions)
+
+    def get_info(self) -> dict:
+        info = super().get_info()
+        info["model_weights"] = self.model_to_weight.copy()
+        return info
+
+    def remap_base_models(self, model_refit_map: dict[str, str]) -> None:
+        updated_weights = {}
+        for model, weight in self.model_to_weight.items():
+            model_full_name = model_refit_map.get(model, model)
+            updated_weights[model_full_name] = weight
+        self.model_to_weight = updated_weights

autogluon/timeseries/models/ensemble/weighted/basic.py

@@ -0,0 +1,91 @@
+from typing import Any
+
+import numpy as np
+
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
+
+from .abstract import AbstractWeightedTimeSeriesEnsembleModel
+
+
+class SimpleAverageEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
+    """Simple ensemble that assigns equal weights to all base models for uniform averaging.
+
+    This ensemble computes predictions as the arithmetic mean of all base model forecasts,
+    giving each model equal influence. Simple averaging is robust and often performs well when base
+    models have similar accuracy levels or when validation data is insufficient to reliably
+    estimate performance differences.
+    """
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ):
+        self.model_to_weight = {}
+        num_models = len(predictions_per_window)
+        for model_name in predictions_per_window.keys():
+            self.model_to_weight[model_name] = 1.0 / num_models
+
+
+class PerformanceWeightedEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
+    """Performance-based weighted ensemble that assigns weights proportional to validation scores.
+
+    This ensemble computes model weights based on their validation performance, giving higher
+    weights to better-performing models. The weighting scheme transforms validation scores
+    (higher is better) into ensemble weights using configurable transformation functions.
+
+    .. warning::
+        This ensemble method is deprecated and may be removed in a future version.
+
+    Other Parameters
+    ----------------
+    weight_scheme : Literal["sq", "inv", "sqrt"], default = "sqrt"
+        Method used to compute the weights as a function of the validation scores.
+
+        - "sqrt" computes weights in proportion to ``sqrt(1 / S)``. This is the default.
+        - "inv" computes weights in proportion to ``(1 / S)``.
+        - "sq" computes the weights in proportion to ``(1 / S)^2`` as outlined in [PC2020]_.
+
+    References
+    ----------
+    .. [PC2020] Pawlikowski, Maciej, and Agata Chorowska.
+        "Weighted ensemble of statistical models." International Journal of Forecasting
+        36.1 (2020): 93-97.
+    """
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        return {"weight_scheme": "sqrt"}
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ):
+        assert model_scores is not None
+
+        weight_scheme = self.get_hyperparameter("weight_scheme")
+
+        # drop NaNs
+        model_scores = {k: v for k, v in model_scores.items() if np.isfinite(v)}
+        assert len(model_scores) > 0, (
+            "All models have NaN scores. At least one model must score successfully to fit an ensemble"
+        )
+        assert all(s <= 0 for s in model_scores.values()), (
+            "All model scores must be negative, in higher-is-better format."
+        )
+
+        score_transform = {
+            "sq": lambda x: np.square(np.reciprocal(x)),
+            "inv": lambda x: np.reciprocal(x),
+            "sqrt": lambda x: np.sqrt(np.reciprocal(x)),
+        }[weight_scheme]
+
+        self.model_to_weight = {
+            model_name: score_transform(-model_scores[model_name] + 1e-5) for model_name in model_scores.keys()
+        }
+        total_weight = sum(self.model_to_weight.values())
+        self.model_to_weight = {k: v / total_weight for k, v in self.model_to_weight.items()}

autogluon/timeseries/models/ensemble/weighted/greedy.py

@@ -0,0 +1,62 @@
+import logging
+import pprint
+from typing import Any
+
+from autogluon.timeseries import TimeSeriesDataFrame
+
+from ..ensemble_selection import fit_time_series_ensemble_selection
+from .abstract import AbstractWeightedTimeSeriesEnsembleModel
+
+logger = logging.getLogger(__name__)
+
+
+class GreedyEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
+    """Greedy ensemble selection algorithm that iteratively builds an ensemble by selecting models with
+    replacement.
+
+    This class implements the Ensemble Selection algorithm by Caruana et al. [Car2004]_, which starts
+    with an empty ensemble and repeatedly adds the model that most improves the ensemble's validation
+    performance. Models can be selected multiple times, allowing the algorithm to assign higher effective
+    weights to better-performing models.
+
+    Other Parameters
+    ----------------
+    ensemble_size : int, default = 100
+        Number of models (with replacement) to include in the ensemble.
+
+    References
+    ----------
+    .. [Car2004] Caruana, Rich, et al. "Ensemble selection from libraries of models."
+        Proceedings of the twenty-first international conference on Machine learning. 2004.
+    """
+
+    def __init__(self, name: str | None = None, **kwargs):
+        if name is None:
+            # FIXME: the name here is kept for backward compatibility. it will be called
+            # GreedyEnsemble in v1.4 once ensemble choices are exposed
+            name = "WeightedEnsemble"
+        super().__init__(name=name, **kwargs)
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        return {"ensemble_size": 100}
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ):
+        model_to_weight = fit_time_series_ensemble_selection(
+            data_per_window=data_per_window,
+            predictions_per_window=predictions_per_window,
+            ensemble_size=self.get_hyperparameter("ensemble_size"),
+            eval_metric=self.eval_metric,
+            prediction_length=self.prediction_length,
+            target=self.target,
+            time_limit=time_limit,
+        )
+        self.model_to_weight = {model: weight for model, weight in model_to_weight.items() if weight > 0}
+
+        weights_for_printing = {model: round(float(weight), 2) for model, weight in self.model_to_weight.items()}
+        logger.info(f"\tEnsemble weights: {pprint.pformat(weights_for_printing, width=200)}")

autogluon/timeseries/models/gluonts/__init__.py

@@ -1,4 +1,4 @@
-from .torch.models import (
+from .models import (
     DeepARModel,
     DLinearModel,
     PatchTSTModel,