capitalone/PRES

Personal and Relational Event Sequence Datasets

This dataset collection supports research on modeling personal and relational event sequences, as referred in our LoG 2025 paper: Integrating Sequential and Relational Modeling for User Events: Datasets and Prediction Tasks. It is organized into two main folders: processed/ and tasks/.

---

Folder Structure

• processed/

  • amazon-clothing_all_events.csv
  • amazon-electronics_all_events.csv
  • brightkite_all_events.csv
  • gowalla_all_events.csv
  • github_all_events.csv

• tasks/

  • amazon-clothing/
    • coreview_prediction/
      • personal_train.csv
      • relational_val.csv
      • relational_test_negative_sample.csv
      • ...
    • product_rating_prediction/
      • personal_val.csv
      • relational_observed.csv
      • ...
  • amazon-electronics/
    • coreview_prediction/
      • ...
    • product_rating_prediction/
      • ...
  • brightkite/
    • checkin_prediction/
      • ...
    • friend_recommendation/
      • ...
  • gowalla/
    • checkin_prediction/
      • ...
    • friend_recommendation/
      • ...
  • github/
    • collab_prediction/
      • ...

---

processed/ Folder

The processed/ folder contains all raw events for each dataset, without any task-specific processing. These files include:

• Full sequences of user events (both personal and relational)
• No train/validation/test split
• No task-specific formatting

Each row includes:

• uid: user ID
• timestamp: event timestamp
• event_set: event set. either "personal" or "relational"
• event: event name
• other_uid: another user ID (for relational events)

---

tasks/ Folder

The tasks/ folder defines various prediction tasks derived from each dataset. Each subfolder corresponds to a dataset and contains directories for specific tasks.

Each task includes:

• Pre-split files for train, val, and test (if applicable)
• Separate files for personal and relational events
• Pre-generated negative samples (e.g., *_negative_sample.csv) for use in evaluation or ranking-based training

---

Notes

• All datasets include both personal and relational event types.
• Task folders follow a uniform schema for easy benchmarking.
• Negative sampling files are provided to support reproducible experiments.

Dataset Description and Sources

pres-brightkite. This dataset contains location check-ins and friendship history of Brightkite users, a location-based social networking platform. It was originally collected by Cho et al., 2011 and published in the SNAP Dataset Repository (Leskovec & Krevl, 2014). Personal events consist of sequences of location check-ins. We convert the original latitude and longitude coordinates into Geohash-8 representations (Niemeyer, 2008; Morton, 1966), short alphanumeric strings encoding geographic locations. Nearby locations share similar geohash prefixes, while distant ones differ. Example geohashes include 9v6kpmr1, gcpwkeq6, and u0yhxgm1. Relational events capture friendship connections among users. The dataset includes 58,228 users and 5,130,866 events. Only personal events have timestamps; relational events do not.

pres-gowalla. The dataset also contains the location check-in and friendship history of another social network platform, Gowalla. It was also originally collected by Cho et al., 2011 and published in the SNAP Repository (Leskovec & Krevl, 2014). We processed and formatted the data following the same approach used for pres-brightkite. The dataset contains personal events from geohash check-ins and relational events from friendship connections, totaling 8,342,943 events from 196,591 users.

pres-amazon-clothing. The dataset contains Amazon product reviews and ratings in the Clothing, Shoes and Jewelry category, spanning from May 1996 to July 2014. The raw data was originally collected by McAuley et al., 2015. In this dataset, we define personal events as sequences of product IDs and ratings reviewed by a user, for example: B000MLDCZ2:5 and B001OE3F08:3. Relational events represent co-review patterns, where two users have reviewed at least three of the same products. The dataset contains event sequences from 185,986 users, with a total of 1,591,947 events.

pres-amazon-electronics. The dataset contains Amazon product reviews and ratings in the Electronics category, originally collected by McAuley et al., 2015. As in pres-amazon-clothing, personal events are defined as sequences of product IDs and ratings, while relational events capture co-review patterns. In total, the dataset contains 2,938,178 events from 254,064 users.

pres-github. This dataset contains GitHub user activity from January 2025, collected from the GH Archive. Personal events include actions such as Push, CreateBranch, CreateRepository, PullRequestOpened, IssuesOpened, and Fork. Relational events represent project collaboration, where two users are linked if both contributed at least five commits or pull requests to the same repository. The dataset includes 102,878,895 events from 3,669,079 users. Only personal events include timestamps; relational events do not, similar to pres-brightkite and pres-gowalla.

References

Cho, E., Myers, S. A., & Leskovec, J. (2011). Friendship and mobility: user movement in location-based social networks. In Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1082–1090.

Leskovec, J., & Krevl, A. (2014). SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data

McAuley, J., Targett, C., Shi, Q., & Van Den Hengel, A. (2015). Image-based recommendations on styles and substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 43–52.

Citation

If you use this dataset, please cite the following paper

@article{fathony2025integrating,
  title={Integrating Sequential and Relational Modeling for User Events: Datasets and Prediction Tasks},
  author={Fathony, Rizal and Melnyk, Igor and Reinert, Owen and Nguyen, Nam H and Rosa, Daniele and Bruss, C Bayan},
  journal={Learning on Graphs Conference 2025},
  year={2025}
}