Research Card on how to make the outcome prediction in online systems fairer?
- Title: FairJob: A Real-World Dataset for Fairness in Online Systems
- Short Title: Discover and mitigate bias in advertising
- Authors: Mariia Vladimirova (Criteo AI Lab), Federico Pavone (Paris-Dauphine), Eustache Diemert (Criteo AI Lab)
- Team: Research.FDL, collaboration with Paris-Dauphine University
- Revue: NeurIPS 2024, datasets and benchmarks track
- Status: Published
- Category: fairness, advertising, tabular data, causal effect
Why did we work on this topic (the problem we want to solve)?
The aim of ad-tech companies is to deliver the most relevant advertisements to users navigating publishers’ webpages. By matching users’ browsing histories and content preferences with products that align with their interests, targeted advertising creates a mutually beneficial ecosystem. Advertisers reach relevant audiences, users have access to free information and services in exchange for seeing ads related to their interests, and platforms profit from selling targeted ads. With the emergence of new technologies like AI, there is a responsibility to ensure that advertising maintains a positive and inclusive impact on society, which includes ensuring fairness in advertising practice.
Ad-tech companies use vast, often historically biased data to predict user behavior. If unchecked, this can worsen biases, harm user satisfaction, damage trust, and even bring legal consequences. For instance, if ads of managerial positions are disproportionately shown to men over women, more men may apply, perpetuating historical biases and exacerbating gender disparities.
How did we proceed?
When the project started, we noticed that to display managerial job opportunities for female gender proxy, Criteo usually pays more than for male gender proxy. It naturally leads to the following questions:
- Where does this bias come from?
- Can we mitigate this bias?
Not only we could answer this question with certainty based on existing research, but there was no even available dataset which (i) would be suited for a fairness problem (having a protected attribute) to test different methods, and (ii) would represent all problems we have with the industrial data (mostly tabular, mixed-type columns, very noisy, strongly unbalanced, large scale).
We did not have much choice and had to proceed with the analysis directly on Criteo’s recommendation data. To further publish our results and promote research on fairness in real-world scenarios, we anonymized the Criteo’s dataset and open-sourced it as the first large-scale fairness-aware dataset for advertising.
What did we find? What did we achieve?
We analyzed at which stages in the job advertising process the bias might be introduced. For example, it was found that women are a prized demographic, making them more expensive to advertise to. This implies that ads that are meant to be gender-neutral can be delivered in the way that appears to be discriminatory by real-time bidding algorithms that focus on optimizing cost-effectiveness. This might affect the future campaign selection through the feedback loop mechanism, resulting into selection bias. And lastly, this might lead to biased data that is used to train the prediction algorithms, thus, inducing algorithmic bias.
The market bias is hard to estimate, as it depends on the entire ecosystem. However, we can estimate the observed selection bias and affect the algorithmic bias in order to reduce the bias in the outcome. Taking into account possible induced biases, we propose an unbiased utility metric that helps to analyse different bias mitigation techniques. We perform experiments of a click prediction on the released dataset to verify how we can improve fairness with respect to gender as a protected attribute.
We compare three methods that employ different training procedures concerning the gender feature:
- Unfair: this method uses all attributes for training, including the protected one — it learns the potential bias directly.
- Unaware: that corresponds to fairness through unawareness. It utilizes all attributes during training except the protected one — it learns the possible bias indirectly through possibly correlated features.
- Fair: that does not use the protected attribute in the data and is trained with an additional fairness-enforcing penalty in the loss.
Our findings in the paper also suggest that the trade-off relationship between accuracy and fairness is context-dependent. It highlights the need for further research to better understand the conditions under which the accuracy fairness trade-off arises and identify strategies to mitigate or overcome it.
What is the originality here?
We became the first to open-source a fairness-aware dataset for advertising, which is also one of the largest fairness-aware tabular dataset coming from a company. Additionally, we described in detail where unfairness can happen during the advertising process which was not presented before in such a detail and comes from our own internal studies. Based on it, we proposed a new metric of utility that takes into account the selection bias and published reproducible results of the experiments on the open-sourced dataset.
Our study demonstrates the possible existence of market bias that might affect the whole ecosystem. We hope our work will help to advance the field of responsible advertising and shed some light on its internal processes.
Check all our research cards 👇
https://medium.com/criteo-engineering/research-cards/home
