Proxy Discrimination: Algorithmic Redlining Lab |…

What Proxy Discrimination Is

Proxy discrimination occurs when an algorithm uses a seemingly neutral feature to indirectly encode information about a protected group. This happens even though the model does not explicitly use any protected attributes, such as race or gender. The 'neutral' proxy can be something like zip code, which may correlate with socioeconomic status and thus indirectly reflect protected characteristics.

This phenomenon is particularly insidious because it can lead to discriminatory outcomes without the algorithm designer's intent, making it a significant challenge in ensuring fair machine learning models.

Why It Happens

The reason proxy discrimination happens lies in the complex relationships between different features. In many real-world datasets, there are strong correlations between seemingly unrelated variables and protected attributes. For example, certain zip codes might be associated with higher or lower average incomes, which can then correlate with other factors like employment rates or creditworthiness.

Machine learning models often rely on these indirect associations to make predictions, leading to outcomes that disproportionately affect specific groups without the model's designers being aware of it.

live demo · related simulation● LIVE

Real-World Examples and Implications

Proxy discrimination has been observed in various domains, including housing, lending, and employment. For instance, a study found that zip codes could be used as a proxy to predict credit risk, leading to higher interest rates for applicants from certain areas. This is problematic because it can perpetuate existing social inequalities without addressing the root causes.

The implications of this phenomenon are far-reaching, affecting not only individual outcomes but also broader societal issues such as economic opportunity and social mobility.

Addressing Proxy Discrimination

To address proxy discrimination, it is crucial to understand the underlying correlations between features in a dataset. Techniques like feature engineering, data preprocessing, and model retraining can help mitigate these issues. However, simply removing protected attributes from the model may not be sufficient, as other features might still encode similar information.

Regulatory frameworks and ethical guidelines are also essential for ensuring that machine learning models do not perpetuate discrimination through proxy features.

Frequently asked questions

Can deleting protected attributes alone fix the issue?

No, simply removing protected attributes from the model does not guarantee fairness. Other features might still encode similar information and lead to discriminatory outcomes.

How can we ensure that machine learning models are fair?

Ensuring fairness requires a multi-faceted approach including understanding feature correlations, using techniques like disparate impact analysis, and implementing ethical guidelines and regulatory frameworks.

What is the four-fifths rule in this context?

The four-fifths rule is a threshold used to determine if there is evidence of disparate treatment or impact. If the approval rate for one group is less than 80% of the approval rate for another group, it may indicate discrimination.

Why does historical bias matter in this lab?

Historical bias matters because it can be encoded in seemingly neutral features like zip codes. This historical data influences model predictions and can perpetuate existing social inequalities without the model designer's intent.

Try it live

Everything above runs in your browser — open Proxy Discrimination: Algorithmic Redlining Lab and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.

▶ Open Proxy Discrimination: Algorithmic Redlining Lab simulation