Let’s not be Hypocrites
My response to the Uber “Dead Miles” Argument
Jay Cassano reached out to me as he was writing his piece on Uber’s “dead miles” (linked below), questioning Uber’s current practice saving driver location data for potential product development. He was specifically interested in a Data Scientist’s opinion on the value of one’s location data — can we put a dollar value to it? And if so, should Uber pay drivers for this data?
A quote from my response to Jay appears in the piece, but my full, slightly-edited email response to Jay is pasted below. While I’m not supportive of Uber’s general data practices, nor am I a fan or their arrogance, the dead miles argument is hard for me to accept.
If we think this is an important battle to fight, then let’s not be hypocrites, and look at practices across the industry as a whole. There are countless services that derive huge value from aggregate user data. While the services themselves may generate revenue, users are not compensated for their portion of data. Users may get free access to a service (e.g. Waze, Facebook), and at other times, they may get access to a service for which they have the opportunity to pay (e.g. Amazon, Netflix, Uber). The aggregate value of this data is huge — Facebook couldn’t generate revenue at scale without it — yet a single user’s dataset is tangential.
How can we estimate $$ value of a single user’s dataset when the real value lies in the aggregate?
Sharing Economy style apps such as Uber cater to two classes of users. For the first, consumers, the application is similar to Netflix and Amazon in the way it supports paid services while learning from aggregate user data. Should we consider drivers any differently? It is true that Uber takes a cut of their fee, but so does Airbnb, and many other digital aggregators (think: Kayak). Why should Uber be treated differently?
The basis of one branch of Machine Learning called supervised learning is to take known (“labeled”) data and train a model that then helps us make accurate predictions on new data. Learning from aggregate data gives us the ability to then make valuable claims about new data that we’ve never seen before. When a new email comes in to our mail server, there’s a spam filter that can make highly accurate inferences whether this message is spam or not. In order to build this type of classifier, we need an initial set of example spam emails, from which we learn and extrapolate. We can continuously train a classifier as new forms of spam come through.
This labeled data, what sits at the core of all supervised machine learning tasks, is critical to the operation of the system. At times we take as input data provided from users of our system. Alternatively, we may use some other dataset that may have been open sourced or made available by a research community. In our industry, t is common practice to utilize low-wage workers to produce this type of labeled data (think Mechanical Turk, CrowdFlower or Grad Students) due to its’ heavily repetitive and manual nature. Some of these labeled data may be made available through open sourced python libraries or other publication venues.
People putting manual labor into creating these labeled data are never compensated based on the potential future revenue that they create.
Is this simply the nature of R&D? Of advancing the state of the art? Rarely are all players involved in the production of an innovative service or product compensated according to their contribution. The startup/VC model is a perfect example of how messy this gets. A combination of luck, good timing, and interpersonal network affects one’s potential monetary gains, rather than impactful contribution to a product.
There’s precedent in other domains.
Rebecca Skloot published a wonderful biography called The Immortal Life of Henrietta Lacks. Henrietta Lacks (and her family) had no idea that her cancerous cells were the basis of the first known human immortal cell line for medical research — critical in advancing Medicine development, specifically helping develop drugs for treating herpes, leukemia, influenza, hemophilia and Parkinson’s disease. Had it not been for the availability of the HeLa cells, we wouldn’t have had as many medical advances.
If services such as Uber, Facebook, Amazon and Google had to charge for data, would they have ever even formed as businesses? Unlikely.
Our actions in aggregate — the labeled datasets that we co-create — serve as the foundation to advances in ML/AI research, and have direct impact on the ability for companies to create innovative services. Some of these services make our lives better, more efficient. Some of the services only serve a privileged section of the population. The more restrictions we make, the more we dampen our industry’s ability to innovate. Are there ways in which we can give users control of their data without limiting technological advancement? Are there financial models which help us compensate users in retrospect, after there’s a clear understanding of the value of an aggregate dataset?
The case of Uber is an interesting one for sure, especially given the way in which they’ve abused their position of power. That said, if we’re going to scrutinize the fact that they’re learning from the routes we take (consumers and drivers alike), we need to scrutinize the industry writ large — Google, Facebook, Amazon, Netflix, and many more companies building services based off of aggregate data. Let’s not be hypocrites.
Curious for your responses.
