Week 10: Related works and figuring out what everyone is talking about

Last week, I mentioned three papers that I had read and summarized for the paper. This week, I read the other two papers and added summaries. Additionally, I took a high-level perspective to the papers regarding fairness and attempted to find the pattern between all their fairness metrics.

Fairness Through Awareness (link)

This paper addresses two different ideas in fairness: group fairness and individual fairness. Group fairness looks at how two different groups are treated and says they should be treated the same, or independently of whatever is being measured. For example, if more men are admitted to college than women (and men and women are equally qualified to attend college), then this violates group fairness. A popular way to calculate group fairness is statistical parity, which just says the outcomes of two groups should not be statistically different. Individual fairness tries to get at the point that no individual should be penalized because of a fairness adjustment. More particularly, the authors say that individual fairness says that two similar individuals should be treated in the same way. For example, if you are in charge of admitting students to college, you might want to admit 50 men and 50 women. You rank the men and women separately, and pick the first 50 applicants from each group.  However, let's say the 51st male applicant is just as qualified as the 50th. Is it fair to not admit that 51st applicant just because you're trying to achieve group fairness?

This paper says that neither pure statistical parity nor pure individual fairness is a good measure, but that both can be approximately achieved using a compromise. Their method starts with perfect statistical parity, and then adjusts it to ensure individuals are treated in the same way. In their method, that 51st applicant would probably be admitted and one of the female applicants would be dropped. They say their method achieves "fair affirmative action".

Identifying Significant Predictive Bias in Classifiers (link)

This paper suggests a way to identify bias within multi-dimensional subgroups. They say that while other papers try to mitigate bias with respect to some known attribute, such as race or gender, these models are not well equipped to consider subgroups with multiple potentially important attributes. Their method uses a scan method that runs in linear time and compares the likelihood ratio score of a subgroup versus all other instances in the data. The model returns the most anomalous subgroup, which you can evaluate to determine if the anomaly is large enough to be considered biased.

What does this all mean?

Good question! I tried to figure that out yesterday by comparing the eight papers about fairness that I have currently read. Here's where I am so far:

There's three big fairness ideas that mainly originated in court cases about unfairness. Those ideas are disparate impact, disparate treatment, and affirmative action. Disparate impact is unintentional bias. For example, admitting students to college based on SAT scores is not outright biased, but SAT scores may be biased towards students from underfunded schools. Disparate treatment is intentional bias. For example, a committee who only admits men because they hate women. Affirmative action is the idea of "fixing" biases by correcting the metric in some way. For example, admitting students to college based on a gender quota.

Several metrics have been developed to evaluate or control for bias at the group level. They work in three ways, which I will call statistical parity, calibration, and equal odds. Statistical parity, as mentioned above, is about having equal outcomes. Calibration is about predicting the right outcome with the same accuracy. (look up "well-calibrated person" for a better idea). Equal odds is about having the same error rates between groups. For example, the false positive rates should be the same between two groups.

Several specific metrics exist in the literature for equal odds, and they go by various names. They are all basically the same, because they compare true positive rate/false negative rate (one can be gained from the other) versus false positive rate/true negative rate (one can be gained from the other). In the literature, these include disparate mistreatment, equalized odds, equal opportunity, balance for the positive class, and balance for the negative class. (see 1, 2, 3, 4, 5, 6, 7, 8)

All these ideas are going to be well summarized with proper citations in the real paper. I'm meeting on Monday with the advisor and with Caitlin, and we will decide proper next steps.

Comments

Post a Comment

Popular posts from this blog

Week 31: Datasets for fairness

This week, Diana and I looked for some datasets on which to evaluate our fairness criteria. This is no simple task. For one, the datasets need to be large enough to be split into training, calibration, and testing sets. Second, there needs to be some protected group attribute like sex, race, or age. Third, there needs to be some outcome ranking as well as some truth value. The truth is often the hardest attribute to find. In this post, I introduce some of the datsets we decided to use. I try to identify each of the necessary features that we need to properly test our fairness criteria. COMPAS The COMPAS dataset is an obvious choice because it is well ingrained in the literature about fairness. It has 18,000 observations (n=18000) and 1000 attributes (k=1000). Protected groups can be sex (81/19) or race (51/49). The outcome attribute can be the recidivism score, the score used in court that represents the risk that the accused will commit another crime. The truth attribute is a li...
Read more

Week 32-33: Wrapping up

Hi all. This is my last blog post for this project. The last few weeks, we spent little time working on the project and instead focused on finishing up our classes. To be honest, we lost some motivation after finishing the paper, and we've been reluctant to start up new tasks so close to the end of the year. In this post, I'd like to talk about some of the things I've learned in this project other than what we found in our research. This post is really more a service to myself; to help me reflect on this past year and what I've learned. When I started this project, I didn't understand how research in computing worked. My exposure to research had been more biology/clinical, in which research involved setting a hypothesis about some protein, designing an experiment with positive and negative controls, and then running the experiment repeatedly until confident in the results. In computing research, my experience has been of two varieties: reading papers and proce...
Read more

Week 30: Returning to Fairness

With the LineUp paper submitted, it's now time to return to the fairness problem we were working on before. As a bit of a refresher for you and for me, I'll review what we've done with fairness so far and what remains to be done. Fairness in Ranking Up until now, the literature regarding fairness has focused on fairness in classification. As an illustrative example, consider fairness in hiring practices with repect to applicant sex. A job opening might receive 100 applicants evenly split into 50 males and 50 females. If males and females are equally capable of doing the job, then you might be concerned to see the hiring company mostly interviewing male applicants. To evaluate fairness, you would compare each applicant's capability (capable of performing duties, not capable) to their outcome (interviewed, not interviewed). In a completely fair scenario, the same proportion of capable applicants from each sex group would receive an interview. Small deviances from this...
Read more