There has been a lot of discussion lately concerning machine learning models that are treating people unfairly. While this isn’t just limited to ML, the speed of development and abstractions that hide the real people impacted by decisions are operating at a scale never experienced before.
What is fairness?
I used to work at Cisco Meraki, a company that started out making cloud managed networking hardware. One of the more interesting problems our firmware teams worked on (I was not involved) was fairly dividing up WIFI capacity among many WIFI devices in a room.
Let us consider the example of a college lecture room with 10 students. Each student has a laptop and a phone. The room has a single WIFI access point (the device that provides WIFI). The 20 student devices all need to share the single access point.
How do we fairly optimize this system? There are two parts to this question: fairness and optimization. How do we define fairness? And how do we then optimize the system?
WIFI is complex because it involves transmitting radio signals through the air. If we are near each other, my transmissions will interfere with your transmissions. To avoid this we either need to be transmitting at different times or at different frequencies.
For our example, the 20 devices are in the same physical space (the room) and are transmitting on the same frequency (that the WIFI radio is using). In order to achieve fairness, the access point must slice up time, giving different blocks of time to different devices.
Each device may have different speeds with which it can transmit and receive. Some devices may be transmitting louder than others (because of physical proximity to the access point). Some students may be playing a video game while others are downloading a document. The gaming students want lower latency while the students downloading documents want better bandwidth. The optimization of this system is complex and highly dynamic as it changes based on what each user is doing, the capability of their devices and the physical environment.
How do we decide to hand out time slices to each device? Do we give each device an equal amount of time? Do we give each device an equal amount of bandwidth, giving slower devices more time? How do we handle bad devices? If the system attempts to give all devices the same bandwidth, a device that operates at 1 KB/s would get 1024x longer than a device that operates at 1 MB/s. Is it fair if a bad device can effectively take down the system?
Fairness is hard
I don’t mean to “solve” this problem. There are many academic papers on this very subject, many of which are implemented in the WIFI standard. I want to highlight that the process of making the world fair is both vague and difficult to achieve.
I recently read “Weapons of Math Destruction” in which the author Cathy O’Neil discusses models that grade job candidates. These models are often entirely statistical; they attempt to judge a candidate based on how that candidate compares to a population. While these models may be “optimized” in that the candidates that they score highly are more likely to do well in the job, the models often “miss” candidates that would do well because the candidate doesn’t fit the standard model. Is this fair?
We have found ourselves in this situation because, in part it is much easier to measure the “optimization” of a model than it is to measure the fairness of a model. “Solving” the WIFI fairness problem is much easier than fairly grading job candidates; there are fewer variables, we can simulate the physics and we can run real-world experiments.
We need to take George Box’s maxim to heart: “All models are wrong, but some are useful”. Be vigilant for ways in which your models might be wrong AND unfair. When evaluating job candidates, I look for reasons to say yes, not for reasons to say no. It is safe to say that automatically grading a person from a score without any human intervention is NOT fair.