A real-world e-commerce dataset for session-based recommender systems research.
A real-world e-commerce dataset for session-based recommender systems research.
The OTTO session dataset is a large-scale, industry-grade dataset designed to bridge the gap between academic research and real-world applications in session-based and sequential recommendation. It features anonymized behavior logs from the OTTO webshop and app, supporting both multi-objective (predicting clicks, carts, and orders) and single-objective tasks. With ready-to-use formats, clear evaluation metrics, and a focus on realistic, scalable research, this dataset aims to drive innovation in the recommender systems community and has been featured in our own Kaggle competition.
clicks, carts and orders.jsonl format| Dataset | #sessions | #items | #events | #clicks | #carts | #orders | Density [%] |
|---|---|---|---|---|---|---|---|
| Train | 12.899.779 | 1.855.603 | 216.716.096 | 194.720.954 | 16.896.191 | 5.098.951 | 0.0005 |
| Test | 1.671.803 | 1.019.357 | 13.851.293 | 12.340.303 | 1.155.698 | 355.292 | 0.0005 |
| mean | std | min | 50% | 75% | 90% | 95% | max | |
|---|---|---|---|---|---|---|---|---|
| Train #events per session | 16.80 | 33.58 | 2 | 6 | 15 | 39 | 68 | 500 |
| Test #events per session | 8.29 | 13.74 | 2 | 4 | 8 | 18 | 28 | 498 |
#events per session histogram (90th percentile)| mean | std | min | 50% | 75% | 90% | 95% | max | |
|---|---|---|---|---|---|---|---|---|
| Train #events per item | 116.79 | 728.85 | 3 | 20 | 56 | 183 | 398 | 129004 |
| Test #events per item | 13.59 | 70.48 | 1 | 3 | 9 | 24 | 46 | 17068 |
#events per item histogram (90th percentile)The data is stored on the Kaggle platform and can be downloaded using their API:
kaggle datasets download -d otto/recsys-datasetThe sessions are stored as JSON objects containing a unique session ID and a list of events:
…session - the unique session idevents - the time ordered sequence of events in the sessionaid - the article id (product code) of the associated eventts - the Unix timestamp of the eventtype - the event type, i.e., whether a product was clicked, added to the user's cart, or ordered during the sessionTo evaluate a model's ability to predict future behavior, as required for deployment in a real-world webshop, we use a time-based validation split. The training set includes user sessions from a 4-week period, while the test set contains sessions from the following week. To prevent information leakage, any training sessions overlapping with the test period were trimmed, ensuring a clear separation between past and future data. The diagram below illustrates this process:
To ensure research relevance and industry applicability, we provide standardized evaluation protocols that closely correlate with real-world performance. For consistent and reliable benchmarking, we strongly recommend:
For click prediction tasks, we recommend using Recall@20 (preferred) and MRR@20, which have demonstrated strong correlation with business impact metrics in our production systems, as validated in our research paper.
| Model | Recall@20 | MRR@20 | Epochs/h |
|---|---|---|---|
| GRU4Rec⁺ | 0.443 | 0.205 | 0.019 |
| SASRec | 0.307 | 0.180 | 0.248 |
| TRON | 0.472 | 0.219 | 0.227 |
For models predicting multiple user actions, we offer two approaches:
Note that multi-objective recommendation evaluation remains an active research area without definitive benchmarks. We welcome further research and contributions to improve evaluation methodologies for these complex scenarios.
For detailed usage instructions and evaluation guidelines regarding the competition, please refer to the KAGGLE.md file.
session defined?aids included in the train set?aids the same as article numbers on otto.de?The OTTO dataset is released under the CC-BY 4.0 License, while the code is licensed under the MIT License.
BibTeX entry:
@online{philipp_normann_sophie_baumeister_timo_wilm_2023,
title={OTTO Recommender Systems Dataset: A real-world e-commerce dataset for session-based recommender systems research},
url={https://www.kaggle.com/dsv/4991874},
doi={10.34740/KAGGLE/DSV/4991874},
publisher={Kaggle},
author={Philipp Normann and Sophie Baumeister and Timo Wilm},
year={2023}
}No open issues yet, or sync has not completed.