MUSES: a benchmark for Marked Unevenly Spaced Event Sequences

MUSES, a benchmark for Marked Unevenly Spaced Event Sequences, is a collection of unevenly spaced time series datasets from various domains, containing marked events for training and evaluating prediction approaches.

Languages

All the columns and classes (when textual) in MUSES are in English (BCP-47 en)

Dataset Structure

Data Instances

All datasets formatting follows a similar structure.

{
    "datetime": [2020-06-10T14:27:37.963000000, 2020-06-20T21:23:06.980000000, 2020-10-15T13:36:16.687000000],
    "type_event": [7, 2, 15],
    "time_since_last_event": [0.0, 246.92472694444444, 2800.2193630555557],
    "time_since_start": [0.0, 246.92472694444444, 3047.14409],
    "seq_idx": 1,
    "dim_process": 50,
    "seq_len": 3
}

earthquake

• Size of downloaded dataset files: 2.77 MB
• Size of the generated dataset: 3.06 MB
• Total amount of disk used: 5.83 MB

memetrack

• Size of downloaded dataset files: 569.01 MB
• Size of the generated dataset: 993.72 MB
• Total amount of disk used: 1562.72 MB

stackoverflow

• Size of downloaded dataset files: 110.71 MB
• Size of the generated dataset: 143.91 MB
• Total amount of disk used: 254.62 MB

taxi_nyc_neighborhoods

• Size of downloaded dataset files: 968.44 MB
• Size of the generated dataset: 2031.69 MB
• Total amount of disk used: 3000.14 MB

synthea

• Size of downloaded dataset files: 7.65 MB
• Size of the generated dataset: 14.56 MB
• Total amount of disk used: 22.21 MB

spiketrains

• Size of downloaded dataset files: 8.14 MB
• Size of the generated dataset: 20.57 MB
• Total amount of disk used: 28.71 MB

crypto_transactions

• Size of downloaded dataset files: 393.35 MB
• Size of the generated dataset: 748.14 MB
• Total amount of disk used: 1141.49 MB

human_activity

• Size of downloaded dataset files: 0.20 MB
• Size of the generated dataset: 0.22 MB
• Total amount of disk used: 0.42 MB

911

• Size of downloaded dataset files: 8.42 MB
• Size of the generated dataset: 21.63 MB
• Total amount of disk used: 30.05 MB

mooc

• Size of downloaded dataset files: 3.97 MB
• Size of the generated dataset: 10.29 MB
• Total amount of disk used: 14.26 MB

amazon_easytpp

• Size of downloaded dataset files: 7.81 MB
• Size of the generated dataset: 10.41 MB
• Total amount of disk used: 18.22 MB

wikipedia

• Size of downloaded dataset files: 127.20 MB
• Size of the generated dataset: 238.24 MB
• Total amount of disk used: 365.43 MB

retweet_easytpp

• Size of downloaded dataset files: 15.12 MB
• Size of the generated dataset: 64.49 MB
• Total amount of disk used: 79.62 MB

taobao_easytpp

• Size of downloaded dataset files: 1.60 MB
• Size of the generated dataset: 2.84 MB
• Total amount of disk used: 4.43 MB

taxi_easytpp

• Size of downloaded dataset files: 0.59 MB
• Size of the generated dataset: 1.88 MB
• Total amount of disk used: 2.47 MB

volcano_easytpp

• Size of downloaded dataset files: 0.15 MB
• Size of the generated dataset: 0.23 MB
• Total amount of disk used: 0.38 MB

hawkes_dependent

• Size of downloaded dataset files: 11.10 MB
• Size of the generated dataset: 15.84 MB
• Total amount of disk used: 26.94 MB

hawkes_1

• Size of downloaded dataset files: 2.04 MB
• Size of the generated dataset: 2.49 MB
• Total amount of disk used: 4.54 MB

Data Fields

The data fields are the same among all datasets and splits. The datetime column is optional and left out when no data is available.

• type_event: a classification label (int64) sequence.
• time_since_last_event: a float64 sequence of inter-event times.
• time_since_start: a float64 sequence of event timestamps since sequence start.
• seq_idx: an int64 feature uniquely identifying sequences in the split.
• dim_process: an int64 feature representing total number of classes in the dataset.
• seq_len: an int64 feature containing sequence length.
• datetime: (if avail.) a timestamp[ns] sequence, containing event datetime if available.

The classification labels differ for each dataset. The corresponding labels for each event type can be read from metadata like this

>>> import datasets as ds
>>> meta_data = ds.load_dataset_builder("ddrg/MUSES", "DATASET-NAME")
>>> [label for label in enumerate(meta_data.info.features["type_event"].feature.names)]

Data Splits

	train	validation	test
earthquake	1724	216	215
memetrack	1264794	158100	158099
stackoverflow	42838	5354	5355
synthea	9198	1150	1149
spiketrains	480	60	60
crypto_transactions	1408527	176066	176066
human_activity	175	22	22
911	1355	169	170
mooc	5638	704	705
amazon_easytpp	7382	922	923
wikipedia	907765	113470	113471
retweet_easytpp	19200	2400	2400
taobao_easytpp	1600	200	200
taxi_easytpp	1600	200	200
volcano_easytpp	333	42	41
hawkes_dependent	19661	2458	2457
hawkes_1	800	100	100

taxi_nyc_neighborhoods

	train	validation	test	extra
taxi_nyc_neighborhoods	273060	34132	34133	106020

Dataset Creation

Curation Rationale

A dataset has to comply with the following hard requirements to be eligible for MUSES.

1. Inherently unevenly spaced
2. Time series minimum length of 2
3. Known license
4. Used in Literature
5. Replicably described data collection and processing
6. Dataset has practical relevance or is commonly used for evaluation
7. Maximum of 100 M events
8. Nominal classes or common class analogy
9. Commonly used in unevenly spaced time or TPP forecasting

Further, we aim to fulfull the following variety constraints by dataset composition.

1. Coverage of many different domains
2. Variety in dataset sizes
3. Variety in sequence lengths
4. Variety in problem difficulties
5. Variety in problem complexity
6. Variety in class balances

See our paper for further information.

Source Data

Initial Data Collection and Normalization

Data was sourced and preprocessed like in other works, unless explicitly stated otherwise in our paper. See Citation Information for details.

Personal and Sensitive Information

Data is deidentified or all personal and sensitive information was removed.

Considerations for Using the Data

Discussion of Biases

There are no ethical biases we are aware of.

Other Known Limitations

Class imbalances are common in event sequences. See the following paper for further information. TODO JIMMY

Additional Information

Dataset Curators

More Information Needed

Licensing Information

The primary MUSES tasks are built on and derived from existing datasets. We refer users to the original licenses accompanying each dataset.

	License	Other
earthquake	Public Domain
memetrack	3-Clause BSD	see LICENSE/memetrack
stackoverflow	CC BY-SA	see LICENSE/stackoverflow
taxi_nyc_neighborhoods	Public Domain
synthea	Apache 2.0	see LICENSE/synthea
spiketrains	CC0	see LICENSE/spiketrains
crypto_transactions	3-Clause BSD	see LICENSE/crypto_transactions
human_activity	CC-BY	see LICENSE/human_activity
911	ODC ODbL	see LICENSE/911
mooc	Scientific Use Only	see LICENSE/mooc
amazon_easytpp	Apache 2.0	see LICENSE/amazon_easytpp
wikipedia	CC BY-SA	see LICENSE/wikipedia
retweet_easytpp	Apache 2.0	see LICENSE/retweet_easytpp
taobao_easytpp	Apache 2.0	see LICENSE/taobao_easytpp
taxi_easytpp	Apache 2.0	see LICENSE/taxi_easytpp
volcano_easytpp	Apache 2.0	see LICENSE/volcano_easytpp
hawkes_dependent	Public Domain
hawkes_1	Public Domain

Citation Information

We encourage you to use the following BibTeX citation for MUSES itself:

# TODO @Jimmy Add here Final Paper

If you use MUSES, please also cite all the individual datasets you use, both to give the original authors their due credit and because venues will expect papers to describe the data they evaluate on. The following provides BibTeX for all of the MUSES tasks.

earthquake

@article{stockman2024earthquakenpp,
  title={EarthquakeNPP: A Benchmark for Earthquake Forecasting with Neural Point Processes},
  author={Stockman, Samuel and Lawson, Daniel and Werner, Maximilian},
  journal={arXiv preprint arXiv:2410.08226},
  year={2024}
}

memetrack

@inproceedings{leskovec2009meme,
  title={Meme-tracking and the dynamics of the news cycle},
  author={Leskovec, Jure and Backstrom, Lars and Kleinberg, Jon},
  booktitle={Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining},
  pages={497--506},
  year={2009}
}
@article{mei2017neural,
  title={The neural hawkes process: A neurally self-modulating multivariate point process},
  author={Mei, Hongyuan and Eisner, Jason M},
  journal={Advances in neural information processing systems},
  volume={30},
  year={2017}
}

stackoverflow

@inproceedings{du2016recurrent,
  title={Recurrent marked temporal point processes: Embedding event history to vector},
  author={Du, Nan and Dai, Hanjun and Trivedi, Rakshit and Upadhyay, Utkarsh and Gomez-Rodriguez, Manuel and Song, Le},
  booktitle={Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining},
  pages={1555--1564},
  year={2016}
}

taxi_nyc_neighborhoods

@inproceedings{du2016recurrent,
  title={Recurrent marked temporal point processes: Embedding event history to vector},
  author={Du, Nan and Dai, Hanjun and Trivedi, Rakshit and Upadhyay, Utkarsh and Gomez-Rodriguez, Manuel and Song, Le},
  booktitle={Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining},
  pages={1555--1564},
  year={2016}
}

synthea

@article{walonoski2018synthea,
  title={Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record},
  author={Walonoski, Jason and Kramer, Mark and Nichols, Joseph and Quina, Andre and Moesel, Chris and Hall, Dylan and Duffett, Carlton and Dube, Kudakwashe and Gallagher, Thomas and McLachlan, Scott},
  journal={Journal of the American Medical Informatics Association},
  volume={25},
  number={3},
  pages={230--238},
  year={2018},
  publisher={Oxford University Press}
}
@inproceedings{enguehard2020neural,
  title={Neural temporal point processes for modelling electronic health records},
  author={Enguehard, Joseph and Busbridge, Dan and Bozson, Adam and Woodcock, Claire and Hammerla, Nils},
  booktitle={Machine Learning for Health},
  pages={85--113},
  year={2020},
  organization={PMLR}
}

spiketrains

@article{stetter2012model,
  title={Model-free reconstruction of excitatory neuronal connectivity from calcium imaging signals},
  author={Stetter, Olav and Battaglia, Demian and Soriano, Jordi and Geisel, Theo},
  year={2012},
  publisher={Public Library of Science San Francisco, USA}
}
@article{linderman2015scalable,
  title={Scalable bayesian inference for excitatory point process networks},
  author={Linderman, Scott W and Adams, Ryan P},
  journal={arXiv preprint arXiv:1507.03228},
  year={2015}
}

crypto_transactions

@article{shamsi2022chartalist,
  title={Chartalist: Labeled graph datasets for utxo and account-based blockchains},
  author={Shamsi, Kiarash and Victor, Friedhelm and Kantarcioglu, Murat and Gel, Yulia and Akcora, Cuneyt G},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={34926--34939},
  year={2022}
}
@inproceedings{du2016recurrent,
  title={Recurrent marked temporal point processes: Embedding event history to vector},
  author={Du, Nan and Dai, Hanjun and Trivedi, Rakshit and Upadhyay, Utkarsh and Gomez-Rodriguez, Manuel and Song, Le},
  booktitle={Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining},
  pages={1555--1564},
  year={2016}
}

human_activity

@article{cook2012casas,
  title={CASAS: A smart home in a box},
  author={Cook, Diane J and Crandall, Aaron S and Thomas, Brian L and Krishnan, Narayanan C},
  journal={Computer},
  volume={46},
  number={7},
  pages={62--69},
  year={2012},
  publisher={IEEE}
}
@article{fortino2021predicting,
  title={Predicting activities of daily living via temporal point processes: Approaches and experimental results},
  author={Fortino, Giancarlo and Guzzo, Antonella and Ianni, Michele and Leotta, Francesco and Mecella, Massimo},
  journal={Computers \& Electrical Engineering},
  volume={96},
  pages={107567},
  year={2021},
  publisher={Elsevier}
}

911

@inproceedings{zuo2020transformer,
  title={Transformer hawkes process},
  author={Zuo, Simiao and Jiang, Haoming and Li, Zichong and Zhao, Tuo and Zha, Hongyuan},
  booktitle={International conference on machine learning},
  pages={11692--11702},
  year={2020},
  organization={PMLR}
}

mooc

@inproceedings{kumar2019predicting,
  title={Predicting dynamic embedding trajectory in temporal interaction networks},
  author={Kumar, Srijan and Zhang, Xikun and Leskovec, Jure},
  booktitle={Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery \& data mining},
  pages={1269--1278},
  year={2019}
}
@article{shchur2019intensity,
  title={Intensity-free learning of temporal point processes},
  author={Shchur, Oleksandr and Bilo{\v{s}}, Marin and G{\"u}nnemann, Stephan},
  journal={arXiv preprint arXiv:1909.12127},
  year={2019}
}

amazon_easytpp

@inproceedings{ni2019justifying,
  title={Justifying recommendations using distantly-labeled reviews and fine-grained aspects},
  author={Ni, Jianmo and Li, Jiacheng and McAuley, Julian},
  booktitle={Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP)},
  pages={188--197},
  year={2019}
}
@article{xue2023easytpp,
  title={Easytpp: Towards open benchmarking temporal point processes},
  author={Xue, Siqiao and Shi, Xiaoming and Chu, Zhixuan and Wang, Yan and Hao, Hongyan and Zhou, Fan and Jiang, Caigao and Pan, Chen and Zhang, James Y and Wen, Qingsong and others},
  journal={arXiv preprint arXiv:2307.08097},
  year={2023}
}

wikipedia

@inproceedings{kumar2019predicting,
  title={Predicting dynamic embedding trajectory in temporal interaction networks},
  author={Kumar, Srijan and Zhang, Xikun and Leskovec, Jure},
  booktitle={Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery \& data mining},
  pages={1269--1278},
  year={2019}
}
@article{bosser2023predictive,
  title={On the predictive accuracy of neural temporal point process models for continuous-time event data},
  author={Bosser, Tanguy and Taieb, Souhaib Ben},
  journal={arXiv preprint arXiv:2306.17066},
  year={2023}
}

retweet_easytpp

@inproceedings{zhao2015seismic,
  title={Seismic: A self-exciting point process model for predicting tweet popularity},
  author={Zhao, Qingyuan and Erdogdu, Murat A and He, Hera Y and Rajaraman, Anand and Leskovec, Jure},
  booktitle={Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining},
  pages={1513--1522},
  year={2015}
}
@article{xue2023easytpp,
  title={Easytpp: Towards open benchmarking temporal point processes},
  author={Xue, Siqiao and Shi, Xiaoming and Chu, Zhixuan and Wang, Yan and Hao, Hongyan and Zhou, Fan and Jiang, Caigao and Pan, Chen and Zhang, James Y and Wen, Qingsong and others},
  journal={arXiv preprint arXiv:2307.08097},
  year={2023}
}

taobao_easytpp

@article{xue2022hypro,
  title={Hypro: A hybridly normalized probabilistic model for long-horizon prediction of event sequences},
  author={Xue, Siqiao and Shi, Xiaoming and Zhang, James and Mei, Hongyuan},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={34641--34650},
  year={2022}
}
@article{xue2023easytpp,
  title={Easytpp: Towards open benchmarking temporal point processes},
  author={Xue, Siqiao and Shi, Xiaoming and Chu, Zhixuan and Wang, Yan and Hao, Hongyan and Zhou, Fan and Jiang, Caigao and Pan, Chen and Zhang, James Y and Wen, Qingsong and others},
  journal={arXiv preprint arXiv:2307.08097},
  year={2023}
}

taxi_easytpp

@article{xue2023easytpp,
  title={Easytpp: Towards open benchmarking temporal point processes},
  author={Xue, Siqiao and Shi, Xiaoming and Chu, Zhixuan and Wang, Yan and Hao, Hongyan and Zhou, Fan and Jiang, Caigao and Pan, Chen and Zhang, James Y and Wen, Qingsong and others},
  journal={arXiv preprint arXiv:2307.08097},
  year={2023}
}

volcano_easytpp

@article{bebbington2014long,
  title={Long-term forecasting of volcanic explosivity},
  author={Bebbington, MS},
  journal={Geophysical Journal International},
  volume={197},
  number={3},
  pages={1500--1515},
  year={2014},
  publisher={Oxford University Press}
}
@misc{easytpp-github-volcano,
  title = {EasyTemporalPointProcess GitHub Volcano Preprocessing},
  author = {Xue, Siqiao and Shi, Xiaoming and Chu, Zhixuan and Wang, Yan and Hao, Hongyan and Zhou, Fan and Jiang, Caigao and Pan, Chen and Zhang, James Y and Wen, Qingsong and others},
  url = {https://github.com/ant-research/EasyTemporalPointProcess/blob/main/examples/script_data_processing/volcano.py}
}

hawkes_dependent

@inproceedings{enguehard2020neural,
  title={Neural temporal point processes for modelling electronic health records},
  author={Enguehard, Joseph and Busbridge, Dan and Bozson, Adam and Woodcock, Claire and Hammerla, Nils},
  booktitle={Machine Learning for Health},
  pages={85--113},
  year={2020},
  organization={PMLR}
}

hawkes_1

@article{omi2019fully,
  title={Fully neural network based model for general temporal point processes},
  author={Omi, Takahiro and Aihara, Kazuyuki and others},
  journal={Advances in neural information processing systems},
  volume={32},
  year={2019}
}

Contributions

Thanks to @Moritz Tschöpe and @Jimmy Pöhlmann for adding this dataset.