This is a summary of excerpts from the article by Cynthia Dwork et al., which investigates the role of fairness in classifying targeted online users. You can get the pdf of the behavioral targeting article here: Fairness Through Awareness.
If an online user responds to an advertisement, it tells the ad network that he or she is interested in the product or service advertised. However, that ad may have been classified by the advertiser as an ad that is directed towards some discriminating classification of users (e.g. rich people who loves whales), which keeps individuals from being treated differently according to their characteristics. It is important that some information should not be used for discriminating individuals. In other words, there should be fairness in classification.
In the future, fairness will definitely play a more important role in the future, thus the motivation of this study. For now, the final test for discrimination will always be upon investigating the end results, but it is advantageous to create a method for designing ads that reduces discrimination.
System Design
Tthe vendor is the external source that provides the classification of individuals. It is imperative that the vendor be given as much freedom so it can gain from data mining, market research, etc. As such, the design for classification should allow the vender to create its own classifier. The system design should therefore be flexible enough to protect this classifier from discrimination for whatever kind of classifier design the vendor creates. Furthermore, the system should adhere to social constraints, including regulations from the government, and protected sets.
Potential yet Insufficient Solutions
These are some of the intuitive solutions to achieving fairness, yet they are incomplete solutions. First is providing fairness by simply not looking at the attributes of consumers. For example, not looking at the sexual preferences of individuals to keep from discriminating against them in these matters. However, studies have shown that these attributes can still be accurately assessed from the other available information. And while these attributes are not know explicitly, they can still discriminate effectively.
Another kind of insufficient solution is by simply randomizing the classifications of individuals. But this clearly reduces the utility of these individuals. For exampe, advertisers would really love to target their customers, so utility and privacy should be balanced.
Design Framework
This study contributes the following methods for achieving fairness. First, fairness is not group-based, but rather individual-based. Therefore, each individual is measured by a distance metric that also measures how similar individuals are from each other. Second, there are two steps in the classification process of individuals. Individuals are grouped together in what are called archetypes, and this is done by the ad network. Then, the advertiser, or what is known as the vendor in this study, will further classify the archetypes into classes.
Third, the framework follows two conditions. One states that individuals who are similar according to the distance metric will be mapped equally. The second condition states that these individuals are classified in a statistical parity, but their individual classifications should reflect the demographics of the whole population. Finaly, it has been observed that the fairness definition generalizes differential privacy, and so how fairness can be used to provide privacy is also discussed.
Credit Card Company Example
An example is a credit card company which classifies their customers through tracking their browsing history, shopping behavior etc. After classification, a person will then be assigned to one of 66 groups, or archetypes. Each group will then be given the kind of credit card that it needs based on the decision of the company. A racial group is also involved and is called the protected set. Because of the statistical parity enforced in this study, members of this set will not be discriminated to be grouped in the least advantageous credit card terms.
