The goal of cesnet-tszoo project is to provide time series datasets with useful tools for preprocessing and reproducibility. Such as:

• API for downloading, configuring and loading various datasets (e.g. CESNET-TimeSeries24, CESNET-AGG23...), each with various sources and aggregations.
• Example of configuration options:
  • Data can be split into train/val/test sets. Split can be done by time series or by time periods.
  • Transforming of data with built-in transformers or with custom transformers.
  • Handling missing values built-in fillers or with custom fillers.
  • Applying custom handlers.
  • Changing order of when are preprocesses applied/fitted
• Creation and import of benchmarks, for easy reproducibility of experiments.
• Creation and import of annotations. Can create annotations for specific time series, specific time or specific time in specific time series.

Datasets

Name	CESNET-TimeSeries24	CESNET-AGG23	Abilene	GÉANT	SDN	Telecom Italia	Network Operator KPIs
Published in	2025	2023	2005	2005	2021	2015	2023
Collection period	9.10.2023 - 14.7.2024	25.2.2023 - 3.5.2023	2004	2005	—	2013–2014	—
Collection duration	40 weeks	10 weeks	6 months	16 weeks	4 days	2 months	Multiple weeks
Aggregation window	1 day, 1 hour, 10 min	1 min	5 min, 10 min, 1 hour, 1 day	15 min, 1 hour, 1 day	1 min, 10 min, 1 hour, 1 day	10 min, 1 hour, 1 day	5 min, 10 min, 1 hour, 1 day
Sources	CESNET3: Institutions, Institution subnets, IP addresses	CESNET2	Abilene network	GÉANT network	Simulated SDN environment	Milan city cells (SMS, call, internet)	Network operator
Subsets	—	—	Matrix, Node2Node, Node	Matrix, Node2Node, Node	Matrix, Node2Node, Node	—	Downstream, Internet, Sessions, VPN
Cite	https://doi.org/10.1038/s41597-025-04603-x	https://doi.org/10.23919/CNSM59352.2023.10327823	https://doi.org/10.1145/885651.781053	https://dl.acm.org/doi/10.1145/1111322.1111341	https://doi.org/10.1109/ICC42927.2021.9500331	https://doi.org/10.1038/sdata.2015.55	https://doi.org/10.5281/zenodo.8147768
Source URL	https://zenodo.org/records/13382427	https://zenodo.org/records/8053021	https://www.cs.utexas.edu/~yzhang/research/AbileneTM	https://totem.info.ucl.ac.be/dataset.html	https://github.com/duchuyle108/SDN-TMprediction	https://dataverse.harvard.edu	https://doi.org/10.5281/zenodo.8147768

Installation

Install the package from pip with:

pip install cesnet-tszoo

or for editable install with:

pip install -e git+https://github.com/CESNET/cesnet-tszoo#egg=cesnet-tszoo

Citation

If you use CESNET TS-Zoo, please cite our paper:

@misc{kures2025,
    title={CESNET TS-Zoo: A Library for Reproducible Analysis of Network Traffic Time Series}, 
    author={Milan Kureš and Josef Koumar and Karel Hynek},
    booktitle={2025 21th International Conference on Network and Service Management (CNSM)}, 
    year={2025}
}

Examples

For detailed examples refer to Tutorial notebooks

Initialize dataset to create train, validation, and test dataframes

Using TimeBasedCesnetDataset dataset

from cesnet_tszoo.datasets import CESNET_TimeSeries24
from cesnet_tszoo.utils.enums import SourceType, AgreggationType, DatasetType
from cesnet_tszoo.configs import TimeBasedConfig

dataset = CESNET_TimeSeries24.get_dataset(data_root="/some_directory/", source_type=SourceType.INSTITUTIONS, aggregation=AgreggationType.AGG_1_DAY, dataset_type=DatasetType.TIME_BASED)
config = TimeBasedConfig(
    ts_ids=50, # number of randomly selected time series from dataset
    train_time_period=range(0, 100), 
    val_time_period=range(100, 150), 
    test_time_period=range(150, 250), 
    features_to_take=["n_flows", "n_packets"])
dataset.set_dataset_config_and_initialize(config)

train_dataframe = dataset.get_train_df()
val_dataframe = dataset.get_val_df()
test_dataframe = dataset.get_test_df()

Time-based datasets are configured with TimeBasedConfig.

Using DisjointTimeBasedCesnetDataset dataset

from cesnet_tszoo.datasets import CESNET_TimeSeries24
from cesnet_tszoo.utils.enums import SourceType, AgreggationType, DatasetType
from cesnet_tszoo.configs import DisjointTimeBasedConfig

dataset = CESNET_TimeSeries24.get_dataset("/some_directory/", source_type=SourceType.INSTITUTIONS, aggregation=AgreggationType.AGG_1_DAY, dataset_type=DatasetType.DISJOINT_TIME_BASED)
config = DisjointTimeBasedConfig(
    train_ts=50, # number of randomly selected time series from dataset that are not in val_ts and test_ts
    val_ts=20, # number of randomly selected time series from dataset that are not in train_ts and test_ts
    test_ts=10, # number of randomly selected time series from dataset that are not in train_ts and val_ts
    train_time_period=range(0, 100), 
    val_time_period=range(100, 150), 
    test_time_period=range(150, 250), 
    features_to_take=["n_flows", "n_packets"])
dataset.set_dataset_config_and_initialize(config)

train_dataframe = dataset.get_train_df()
val_dataframe = dataset.get_val_df()
test_dataframe = dataset.get_test_df()

Disjoint-time-based datasets are configured with DisjointTimeBasedConfig.

Using SeriesBasedCesnetDataset dataset

from cesnet_tszoo.datasets import CESNET_TimeSeries24
from cesnet_tszoo.utils.enums import SourceType, AgreggationType, DatasetType
from cesnet_tszoo.configs import SeriesBasedConfig

dataset = CESNET_TimeSeries24.get_dataset(data_root="/some_directory/", source_type=SourceType.INSTITUTIONS, aggregation=AgreggationType.AGG_1_DAY, dataset_type=DatasetType.SERIES_BASED)
config = SeriesBasedConfig(
    time_period=range(0, 250), 
    train_ts=50, # number of randomly selected time series from dataset that are not in val_ts and test_ts
    val_ts=20, # number of randomly selected time series from dataset that are not in train_ts and test_ts
    test_ts=10, # number of randomly selected time series from dataset that are not in train_ts and val_ts
    features_to_take=["n_flows", "n_packets"])
dataset.set_dataset_config_and_initialize(config)

train_dataframe = dataset.get_train_df()
val_dataframe = dataset.get_val_df()
test_dataframe = dataset.get_test_df()

Series-based datasets are configured with SeriesBasedConfig.

Using load_benchmark

from cesnet_tszoo.benchmarks import load_benchmark

benchmark = load_benchmark(identifier="2e92831cb502", data_root="/some_directory/")
dataset = benchmark.get_initialized_dataset()

train_dataframe = dataset.get_train_df()
val_dataframe = dataset.get_val_df()
test_dataframe = dataset.get_test_df()

Loaded dataset can be one of the above.