Forecaster

HyperTreeNetARForecaster

Categorical in XPred int insampleExogenous

Hyper-TreeNet-AR forecaster, from the hypertrees-forecasting package.

Direct interface to hypertrees.models.HyperTreeNetAR [1].

Hyper-Trees use a gradient boosted tree (LightGBM) to learn the parameters of a classical time series model as functions of features, rather than forecasting the series directly. HyperTreeNetAR targets a time-varying AR(p) model: the tree produces embeddings that a small neural network maps to the AR parameters, and the AR model generates the forecast.

The interfaced estimator is univariate and models a single series.

Quickstart

python

from sktime.forecasting.hypertrees import HyperTreeNetARForecaster

estimator = HyperTreeNetARForecaster(p=2, embedding_dimension=1, hidden_dim=128, dropout=0.1, use_random_projection=True, rp_embed_dim=12, network_learning_rate=0.001, gradient_mode='separate', device='cpu', hessian_method='exact', n_hessian_probes=5, lgb_params=None, num_iterations=100, seed=123)

Parameters(14)

pint, optional (default=2): Maximum number of AR(p) lags.
embedding_dimensionint, optional (default=1): Embedding dimension of the tree embeddings fed to the network.
hidden_dimint, optional (default=128): Hidden dimension of the embedding network (MLP).
dropoutfloat, optional (default=0.1): Dropout rate of the embedding network.
use_random_projectionbool, optional (default=True): Whether to use random projections for the embeddings.
rp_embed_dimint, optional (default=12): Dimension of the random projections, only used when use_random_projection=True.
network_learning_ratefloat, optional (default=1e-3): Learning rate of the embedding network optimizer.
gradient_modestr, optional (default=”separate”): Gradient computation mode, "separate" or "shared".
devicestr, optional (default=”cpu”): Device for the embedding network, e.g. "cpu" or "cuda".
hessian_methodstr, optional (default=”exact”): Method for the Hessian diagonal, "exact" or "gn".
n_hessian_probesint, optional (default=5): Number of Hutchinson probes, only used when hessian_method="gn".
lgb_paramsdict, optional (default=None): LightGBM parameters. If None, {"learning_rate": 0.1} is used.
num_iterationsint, optional (default=100): Number of boosting rounds.
seedint, optional (default=123): Random seed for the interfaced estimator.

Examples

>>> from sktime.forecasting.hypertrees import HyperTreeNetARForecaster
>>> from sktime.datasets import load_airline
>>> y = load_airline ()
>>> forecaster = HyperTreeNetARForecaster (p = 2)
>>> forecaster. fit (y, fh = [1, 2, 3 ]) HyperTreeNetARForecaster(
... )
>>> y_pred = forecaster. predict ()

References

[1]

Maerz, Alexander, and Kashif Rasul. “Forecasting with Hyper-Trees.” arXiv preprint arXiv:2405.07836 (2024).