FairJob: A Real-World Dataset for Fairness in Online Systems

Summary

This dataset is released by Criteo to foster research and innovation on Fairness in Advertising and AI systems in general. See also Criteo pledge for Fairness in Advertising.

The dataset is intended to learn click predictions models and evaluate by how much their predictions are biased between different gender groups. The associated paper is available at Vladimirova et al. 2024.

License

The data is released under the CC-BY-NC-SA 4.0 license. You are free to Share and Adapt this data provided that you respect the Attribution, NonCommercial and ShareAlike conditions. Please read carefully the full license before using.

Data description

The dataset contains pseudononymized users' context and publisher features that was collected from a job targeting campaign ran for 5 months by Criteo AdTech company. Each line represents a product that was shown to a user. Each user has an impression session where they can see several products at the same time. Each product can be clicked or not clicked by the user. The dataset consists of 1072226 rows and 55 columns.

• features
  • user_id is a unique identifier assigned to each user. This identifier has been anonymized and does not contain any information related to the real users.
  • product_id is a unique identifier assigned to each product, i.e. job offer.
  • impression_id is a unique identifier assigned to each impression, i.e. online session that can have several products at the same time.
  • cat0 to cat5 are anonymized categorical user features.
  • cat6 to cat12 are anonymized categorical product features.
  • num13 to num47 are anonymized numerical user features.
• labels
  • protected_attribute is a binary feature that describes user gender proxy, i.e. female is 0, male is 1. The detailed description on the meaning can be found below.
  • senior is a binary feature that describes the seniority of the job position, i.e. an assistant role is 0, a managerial role is 1. This feature was created during data processing step from the product title feature: if the product title contains words describing managerial role (e.g. 'president', 'ceo', and others), it is assigned to 1, otherwise to 0.
  • rank is a numerical feature that corresponds to the positional rank of the product on the display for given impression_id. Usually, the position on the display creates the bias with respect to the click: lower rank means higher position of the product on the display.
  • displayrandom is a binary feature that equals 1 if the display position on the banner of the products associated with the same impression_id was randomized. The click-rank metric should be computed on displayrandom = 1 to avoid positional bias.
  • click is a binary feature that equals 1 if the product product_id in the impression impression_id was clicked by the user user_id.

Data statistics

dimension	average
click	0.077
protected attribute	0.500
senior	0.704

Protected attribute

As Criteo does not have access to user demographics we report a proxy of gender as protected attribute. This proxy is reported as binary for simplicity yet we acknowledge gender is not necessarily binary.

The value of the proxy is computed as the majority of gender attributes of products seen in the user timeline. Product having a gender attribute are typically fashion and clothing. We acknowledge that this proxy does not necessarily represent how users relate to a given gender yet we believe it to be a realistic approximation for research purposes.

We encourage research in Fairness defined with respect to other attributes as well.

Limitations and interpretations

We remark that the proposed gender proxy does not give a definition of the gender. Since we do not have access to the sensitive information, this is the best solution we have identified at this stage to idenitify bias on pseudonymised data, and we encourage any discussion on better approximations. This proxy is reported as binary for simplicity yet we acknowledge gender is not necessarily binary. Although our research focuses on gender, this should not diminish the importance of investigating other types of algorithmic discrimination. While this dataset provides important application of fairness-aware algorithms in a high-risk domain, there are several fundamental limitation that can not be addressed easily through data collection or curation processes. These limitations include historical bias that affect a positive outcome for a given user, as well as the impossibility to verify how close the gender-proxy is to the real gender value. Additionally, there might be bias due to the market unfairness. Such limitations and possible ethical concerns about the task should be taken into account while drawing conclusions from the research using this dataset. Readers should not interpret summary statistics of this dataset as ground truth but rather as characteristics of the dataset only.

Challenges

The first challenge comes from handling the different types of data that are common in tables, the mixed-type columns: there are both numerical and categorical features that have to be embedded. In addition, some of the features have long-tail phenomenon and products have popularity bias. Our datasets contains more than 1,000,000 lines, while current high-performing models are under-explored in scale. Additional challenge comes from strongly imbalanced data: the positive class proportion in our data is less than 0.007 that leads to challenges in training robust and fair machine learning models. In our dataset there is no significant imbalances in demographic groups users regarding the protected attribute (both genders are sub-sampled with 0.5 proportion, female profile users were shown less job ad with 0.4 proportion and slightly less senior position jobs with 0.48 proportion), however, there could be a hidden effect of a selection bias. This poses a problem in accurately assessing model performance. More detailed statistics and exploratory analysis are referred to the supplemental material of the associated paper linked below.

Metrics

We strongly recommend to measure prediction quality using Negative Log-likelihood (lower is better).

We recommend to measure Fairness of ads by Demographic Parity conditioned on Senior job offers:

$$E[f(x)\mathrm{\mid }protected\mathrm{\_}attribute=1,senior=1]-E[f(x)\mathrm{\mid }protected\mathrm{\_}attribute=0,senior=1]E[f(x)\; |\; protected\backslash \_attribute=1,\; senior=1]\; -\; E[f(x)\; |\; protected\backslash \_attribute=0,\; senior=1]$$

This corresponds to the average difference in predictions for senior job opportunities between the two gender groups (lower is better). Intuitively, when this metric is low it means we are not biased towards presenting more senior job opportunities (e.g. Director of XXX) to one gender vs the other.

Example

You can start by running the example in example.py (requires numpy + torch). This implements

• a dummy classifier (totally fair yet not very useful)
• a logistic regression with embeddings for categorical features (largely unfair and useful)
• a "fair" logistic regression (relatively fair and useful)

The "fair" logistic regression is based on the method proposed by Bechavod et al. 2017.

Citation

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

@misc{vladimirova2024fairjob,
      title={FairJob: A Real-World Dataset for Fairness in Online Systems}, 
      author={Mariia Vladimirova and Federico Pavone and Eustache Diemert},
      year={2024},
      eprint={2407.03059},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2407.03059},
}