Pyrregular: Irregular Time Series Datasets and Benchmarks

Published at ICLR 2026 · benchmark of 34 datasets and 12 classifiers · pip install pyrregular

Pyrregular is a Python framework for irregular time series: irregularly sampled data with uneven sampling, missing observations, signals recorded at different times, and variable-length sequences. It provides:

• datasets: the 34 naturally irregular datasets of the ICLR 2026 benchmark, and more added over time, in one standardized format, downloaded on demand with load_dataset (list_datasets() shows them all);
• a representation that keeps the irregularity: a sparse xarray format, with conversions to dense arrays for sktime, aeon, tslearn and PyPOTS;
• models and benchmarks: ready-to-use classification and regression pipelines, including the 12 classifiers evaluated in the paper.

Irregular time series made easy.

	📖 Documentation · ⚙️ Tutorials
CI/CD
Code
Community
Paper

Installation

You can install via pip with:

pip install pyrregular

For the third-party models, install all of them:

pip install "pyrregular[models]"

or only the family you need: borf, sktime, tslearn, pypots, jax (e.g. pip install "pyrregular[sktime]"). The extra for each pipeline is listed in the table below.

Quick Guide

List datasets

If you want to see all the datasets available, you can use the list_datasets function:

from pyrregular import list_datasets

df = list_datasets()

Load a dataset

To load a dataset, you can use the load_dataset function. For example, to load the "Garment" dataset, you can do:

from pyrregular import load_dataset

df = load_dataset("Garment.h5")

The dataset is saved in the default os cache directory, which can be found with:

import pooch

print(pooch.os_cache("pyrregular"))

The repository is hosted at: https://huggingface.co/datasets/splandi/pyrregular/

The datasets are versioned: each pyrregular release downloads a fixed version of the files, tagged on Hugging Face (currently data-v1). To use the data of an older release, install that release (e.g. pip install "pyrregular==0.3.1").

Downstream tasks

Classification

To use the dataset for classification, you can just "densify" it:

from pyrregular import load_dataset

df = load_dataset("Garment.h5")
X, _ = df.irr.to_dense()
y, split = df.irr.get_task_target_and_split()

X_train, X_test = X[split != "test"], X[split == "test"]
y_train, y_test = y[split != "test"], y[split == "test"]

# We have ready-to-go models from various libraries:
from pyrregular.models.rocket import rocket_pipeline

model = rocket_pipeline
model.fit(X_train, y_train)
model.score(X_test, y_test)

There are several pipelines available in pyrregular.models:

💾 Library	📖 Source	🔗 Pipeline	ℹ️ Type	📦 Extra
fast-borf	Spinnato et al. (2024)	borf	dictionary-based transform + lgbm classifier	borf
fast-borf	Spinnato (2026)	iborf	time-aware dictionary-based transform + ridge classifier	borf
sktime		rifc	interval-based transform + lgbm classifier	sktime
diffrax	Kidger et al. (2020)	ncde	neural controlled differential equations	jax
pypots	Cao et al. (2018)	brits	bidirectional recurrent imputation network	pypots
pypots	Che et al. (2018)	grud	gated recurrent unit with decay	pypots
pypots	Zhang et al. (2021)	raindrop	graph neural network	pypots
pypots	Du et al. (2023)	saits	self-attention-based imputation transformer	pypots
pypots	Wu et al. (2022)	timesnet	temporal 2d-variation transformer	pypots
sktime	Ke et al. (2017)	lgbm	gradient boosted tree	sktime
sktime	Dempster et al. (2021)	rocket	kernel-based transform + lgbm classifier	sktime
sktime	Bagheri et al. (2016)	svm	support vector machine with distance kernel	sktime
tslearn	Sakoe & Chiba (1978)	knn	distance-based with dynamic time warping	tslearn

Notes

• iborf is in pyrregular.models.borf (from pyrregular.models.borf import iborf_pipeline) and uses the timestamps: give it X, _ = df.irr.to_dense(concatenate_time=True, normalize_time=True).
• raindrop also needs torch-scatter, built for your torch/CUDA version (see the PyG install instructions).
• To reproduce the paper, list_paper_models() lists its 12 models and pyrregular.models.get_paper_model(name) returns a fresh copy of each pipeline. The paper's BORF is the aeon implementation (pip install "pyrregular[aeon]").
• These pipelines wrap fast-moving third-party libraries. They are tested weekly against the latest versions, but an upstream release can break one between pyrregular releases; to reproduce the paper results exactly, pin the library versions.

Regression

Regression is still work in progress, but is available for some datasets:

from pyrregular import load_dataset
from sklearn.pipeline import make_pipeline
from aeon.transformations.collection.dictionary_based import BORF
from sklearn.linear_model import LassoCV

df = load_dataset("Garment.h5")
X, _ = df.irr.to_dense()
y, split = df.irr.get_task_target_and_split("regression")

X_train, X_test = X[split != "test"], X[split == "test"]
y_train, y_test = y[split != "test"], y[split == "test"]

borf_pipeline = make_pipeline(
    BORF(),
    LassoCV(),
)
model = borf_pipeline
model.fit(X_train, y_train)
model.score(X_test, y_test)

Available Datasets

📈 Dataset	📖 Source
Alembics Bowls Flasks	Spinnato & Landi, 2025
AllGestureWiimoteX	Guna et al., 2014
AllGestureWiimoteY	Guna et al., 2014
AllGestureWiimoteZ	Guna et al., 2014
Animals	Ferrero et al., 2018
AsphaltObstaclesCoordinates	Souza, 2018
AsphaltPavementTypeCoordinates	Souza, 2018
AsphaltRegularityCoordinates	Souza, 2018
CharacterTrajectories	Williams et al., 2006
DodgerLoopDay	Ihler et al., 2006
DodgerLoopGame	Ihler et al., 2006
DodgerLoopWeekend	Ihler et al., 2006
Geolife	Zheng et al., 2009; Zheng et al., 2008; Zheng et al., 2010
GestureMidAirD1	Caputo et al., 2018
GestureMidAirD2	Caputo et al., 2018
GestureMidAirD3	Caputo et al., 2018
GesturePebbleZ1	Mezari & Maglogiannis, 2018
GesturePebbleZ2	Mezari & Maglogiannis, 2018
GPS Data of Seabirds	Browning et al., 2018
InsectWingbeat	Chen et al., 2014
JapaneseVowels	Kudo et al., 1999
Localization Data for Person Activity	Vidulin et al., 2010
MelbournePedestrian	City of Melbourne, 2019
MIMIC-III Clinical Database (Demo)	Johnson et al., 2016; Johnson et al., 2019; Goldberger et al., 2000
PAMAP2 Physical Activity Monitoring	Reiss & Stricker, 2012
PhysioNet 2012	Silva et al., 2012
PhysioNet 2019	Reyna et al., 2020
PickupGestureWiimoteZ	Guna et al., 2014
PLAID	Gao et al., 2014
Productivity Prediction of Garment Employees	Imran et al., 2021
ShakeGestureWiimoteZ	Guna et al., 2014
SpokenArabicDigits	Hammami & Bedda, 2010
Taxi	Moreira-Matias et al., 2013
Vehicles	Chorochronos Archive, 2019

Citation

If you use this package in your research, please cite the following paper:

@inproceedings{
    spinnato2026pyrregular,
    title={{PYRREGULAR}: A Unified Framework for Irregular Time Series, with Classification Benchmarks},
    author={Francesco Spinnato and Cristiano Landi},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=qetBM8nLkf}
}