A real-world e-commerce dataset for session-based recommender systems research.
Get the Data • Data Format • Installation • Evaluation • FAQ • License
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.
- 12M real-world anonymized user sessions
- 220M events, consiting of
clicks,cartsandorders - 1.8M unique articles in the catalogue
- Ready to use data in
.jsonlformat - Evaluation metrics for single and multi-objective tasks
| 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": 42,
"events": [
{ "aid": 0, "ts": 1661200010000, "type": "clicks" },
{ "aid": 1, "ts": 1661200020000, "type": "clicks" },
{ "aid": 2, "ts": 1661200030000, "type": "clicks" },
{ "aid": 2, "ts": 1661200040000, "type": "carts" },
{ "aid": 3, "ts": 1661200050000, "type": "clicks" },
{ "aid": 3, "ts": 1661200060000, "type": "carts" },
{ "aid": 4, "ts": 1661200070000, "type": "clicks" },
{ "aid": 2, "ts": 1661200080000, "type": "orders" },
{ "aid": 3, "ts": 1661200080000, "type": "orders" }
]
}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 session
To 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:
- Using the provided train/test split to ensure direct comparability with other research results, without leaving out any items or sessions in the evaluation
- Evaluating on the entire test sequences without truncation
- Never use sampling during evaluation, as this will lead to misleading results (see details here)
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:
- Joint Recall Metric: Developed for our Kaggle competition, this metric integrates recall scores for clicks, basket additions, and orders into a single comprehensive measure
- MultiTRON: An approach that optimizes for clicks and orders simultaneously, allowing for evaluation of different preference trade-offs as detailed in our research paper
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.
- A session is all activity by a single user either in the train or the test set.
- No, train and test users are completely disjunct.
- Yes, all test items are also included in the train set.
- This can happen if the ordered item was already in the customer's cart before the data extraction period started. Similarly, a wishlist in our shop can lead to cart additions without a previous click.
Are aids the same as article numbers on otto.de?
- No, all article and session IDs are anonymized.
- No, our current recommendations generated only about 20% of the product page views in the dataset. Most users reached product pages via search results and product lists.
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}
}