README.md

---
license: cc-by-sa-4.0
size_categories:
- 10M<n<100M
task_categories:
- tabular-classification
- tabular-regression
tags:
- criteo
- advertising
---
# Dataset Documentation

## Private Bidding Optimisation {#private-conversion-optimisation}

The advertising industry lacks a common benchmark to assess the privacy
/ utility trade-off in private advertising systems. To fill this gap, we
are open-sourcing CriteoPrivateAd, the largest real-world anonymised
bidding dataset, in terms of number of features. This dataset enables
engineers and researchers to:

-   assess the impact of removing cross-domain user signals,
    highlighting the effects of third-party cookie deprecation;
-   design and test private bidding optimisation approaches using
    contextual signals and user features;
-   evaluate the relevancy of answers provided by aggregation APIs for
    bidding model learning.

## Summary

This dataset is released by Criteo to foster research and industrial
innovation on privacy-preserving machine learning applied to a major
advertising use-case, namely bid optimisation under user signal loss /
obfuscation.

This use-case is inspired by challenges both browser vendors and AdTech
companies are facing due to third-party cookie deprecation, such as
ensuring a viable cookie-less advertising business via a pragmatic
performance / privacy trade-off. In particular, we are expecting to see
improvements of Google Chrome Privacy Sandbox and Microsoft Ad Selection
APIs via offline benchmarks based on this dataset.

The dataset contains an anonymised log aiming to mimic production
performance of AdTech bidding engines, so that offline results based on
this dataset could be taken as ground truth to improve online
advertising performance under privacy constraints. Features are grouped
into several groups depending on their nature, envisioned privacy
constraints and availability at inference time.

Based on this dataset, the intended objective is to implement privacy
constraints (e.g. by aggregating labels or by adding differential
privacy to features and/or labels) and then learn click and conversion
(e.g. sales) prediction models.

The associated paper is available [here](https://arxiv.org/abs/2502.12103)

As a leading AdTech company that drives commerce outcomes for media
owners and marketers, Criteo is committed to evaluating proposals that
might affect the way we will perform attribution, reporting and campaign
optimisation in the future. Criteo has already participated in testing
and providing feedback on browser proposals such as the Privacy Sandbox
one; see all our [Medium articles](https://techblog.criteo.com) Back in 2021, we also
organised a public challenge aiming to assess bidding performance when
learning on aggregated data: our learnings are available [here](https://arxiv.org/abs/2201.13123).



## Dataset Description

A precise description of the dataset and each column is available in [the
companion paper](https://arxiv.org/abs/2502.12103)

This dataset represents a 100M anonymised sample of 30 days of Criteo
live data retrieved from third-party cookie traffic on Chrome. Each line corresponds to one impression (a banner)
that was displayed to a user. It is partionned by day (`day_int`) to facilitate exploration, model seeding and train/validation/test split.

For each impression, we are providing:

- campaign x publisher x (user x day) granularity with respective ids, to match Chrome Privacy Sandbox scenarios and both
display and user-level privacy.

-   4 labels (click, click leading to a landing on an advertiser
    website, click leading to a visit on an advertiser website -
    i.e. landing followed by one advertiser event, number of sales
    attributed to the clicked display).

-   more than 100 features grouped in 5 buckets with respect to their
    logging and inference constraints in Protected Audience API from
    Chrome Privacy Sandbox (note that these buckets are generic enough
    to cover other private advertising frameworks as we are mainly
    providing a split between ad campaign features, single-domain &
    cross-domain user features, and contextual features) :

    -   Features available in the key-value server with 12-bit logging
        constraint (i.e. derived from current version of modelingSignals
        and standing for single-domain user features).
    -   Features available in the key-value server with no logging
        constraint (i.e. derived from Interest Group name / renderURL).
    -   Features available in browser with 12-bit constraint
        (i.e. cross-domain features available in generateBid).
    -   Features from contextual call with no logging constraint
        (i.e. contextual features).
    -   Features not available (i.e. cross-device and cross-domain
        ones).

-   `day_int` enabling (1) splitting the log into training, validation
    and testing sets; (2) performing relevant model seeding.

-   Information about conversion delay to simulate the way Privacy Sandbox APIs are working.

-   `time_between_request_timestamp_and_post_display_event` (column name
    in clear): time delta (in minutes) between the request timestamp and the
    click or sale event. All displays are considered starting the day of
    the event at 00:00 to avoid providing complete timelines.

-   We include a display order from 1 to K for display on the same day
    for the same user.

The displays-per-user histograms can be deduced from event_per_user_contribution.csv. It is useful to build importance sampling ratios and user-level DP, as it is detailed in the companion paper.






## Metrics

The metrics best suited to the click and conversion estimation problems
are:

-   the log-likelihood (LLH), and preferably a rescaled version named LLH-CompVN defined
    as the relative log-likelihood uplift compared to the naive model
    always predicting the average label in the training dataset;
-   calibration, defined as the ratio between the sum of the predictions
    and the sum of the validation labels. It must be close to 1 for a
    bidding application;

We would like to point out that conventional classification measures
such as area under the curve (AUC) are less relevant for comparing
auction models.

The click-through rate is higher than the one encountered in real-world
advertising systems on the open internet. To design realistic bidding
applications, one must use a weighted loss for validation. We defer the
interested readers to the [associated companion paper](https://arxiv.org/abs/2502.12103) for more details

## Baselines

The Training period has been fixed to 1->25 and Validation period to 26->30. The chosen loss is the LLH-CompVN with weighting as defined above. The Sales | Display is a product of the Landed Click | Display and the Sales | Landed Click. 

| Task/CTR                | 0.1%  | 0.5%  | 1%    |
|-------------------------|-------|-------|-------|
| Landed Click \| Display  | 0.170 | 0.186 | 0.234 |
| Sales \| Landed Click   | 0.218 | 0.218 | 0.218 |
| Sales \| Display        | 0.171 | 0.187 | 0.237 |

Note that our baseline results might be difficult to achieve because of the anonymisation of the dataset.

## License

The data is released under the  license. You are free to
Share and Adapt this data provided that you respect the Attribution and
ShareAlike conditions. Please read carefully the full license before
using.

## Citation

If you use the dataset in your research please cite it using the
following Bibtex excerpt:

```bibtex
@misc{sebbar2025criteoprivateadrealworldbiddingdataset,
  title={CriteoPrivateAd: A Real-World Bidding Dataset to Design Private Advertising Systems},
  author={Mehdi Sebbar and Corentin Odic and Mathieu Léchine and Aloïs Bissuel and 
          Nicolas Chrysanthos and Anthony D'Amato and Alexandre Gilotte and 
          Fabian Höring and Sarah Nogueira and Maxime Vono},
  year={2025},
  eprint={2502.12103},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2502.12103},
}
```


## Acknowledgment

We would like to thank:

-   Google Chrome Privacy Sandbox team, especially Charlie Harrison,
    for feedbacks on the usefulness of this dataset.
-   W3C PATCG group, notably for their public data requests to foster
    work on the future of attribution and reporting.
-   Criteo stakeholders who took part of this dataset release: Anthony
    D'Amato, Mathieu Léchine, Mehdi Sebbar, Corentin Odic, Maxime Vono,
    Camille Jandot, Fatma Moalla, Nicolas Chrysanthos, Romain Lerallut,
    Alexandre Gilotte, Aloïs Bissuel, Lionel Basdevant, Henry Jantet.