Big Data, Plainly Spoken (aka Numbers Rule Your World)

Data sleaze: Uber and beyond

There has been a barrage of negative publicity related to Uber recently. The latest salvo is a long article in the New York Times (link). This piece focuses on Uber's CEO, who was trained as a computer engineer, but my interest lies primarily in several revelations about how Uber collects and uses customer data.

The key episode picked up by various outlets (e.g. TechCrunch, Wired) involves Uber "secretly identifying and tagging iPhones even after its app had been deleted and the devices erased." What Uber engineers did was against Apple rules, and they knew it because they also implemented an elaborate cover-up operation but Apple eventually discovered the ruse, and CEO Tim Cook called Uber to task.

Much of the reporting on this episode, whether in the Times or Wired, etc., miss the mark. These reporters seemed to have received Uber's side of the story from their PR team, and just printed it without asking the tough questions. Uber claimed that their rule-breaking code was intended to "prevent fraud," and suggested that it is a standard practice used by many other app developers.

If those assertions were true, then Apple would have summoned a lot of developers to Cupertino! Uber clearly went beyond what other app developers were doing.

Also, it is not at all clear how this alleged fraud detection scheme works. Here are the details offered by the Times:

At the time, Uber was dealing with widespread account fraud in places like China, where tricksters bought stolen iPhones that were erased and resold. Some Uber drivers there would then create dozens of fake email addresses to sign up for new Uber rider accounts attached to each phone, and request rides from those phones, which they would then accept. Since Uber was handing out incentives to drivers to take more rides, the drivers could earn more money this way.

To halt the activity, Uber engineers assigned a persistent identity to iPhones with a small piece of code, a practice called “fingerprinting.” Uber could then identify an iPhone and prevent itself from being fooled even after the device was erased of its contents.

As someone with an engineering degree, I don't understand what those words mean. First, it seems that erasing the device does not remove the Uber-added piece of code, which is perhaps more of an uncomfortable question for Apple than for Uber. Second, am I to believe that there are no legitimate, refurbished iPhones in use? That last sentence sounds like every iPhone that switched users is a fraud. Third, there are many other ways to detect fraud - those fake user accounts have to be tied to credit cards, for example. Fourth, if a promotion encourages gaming and fraud, to the extent that a major software development effort, including a cover-up operation, is needed to support it, maybe the promotion should be retired, no?

***

Readers should definitely read behind the lines here. The most noteworthy items weren't even mentioned.

First, Uber is after personally identifying data. If we believe that fraud detection was the original intent of such unruly fingerprinting, then Uber wants to know which customer is using a specific phone, which implies that the data being collected must identify individuals (as well as individual phones).

Second, while Uber claimed that the code was developed for fraud detection, nowhere did it deny that the collected data have been used for other purposes, such as marketing. Such data is extremely useful for acquiring new customers. It tells Uber which phones do not have their app installed. It is as if the first owner of the iPhone who installs the Uber app places every future owner of that iPhone onto Uber's prospect list.

The data is also extremely usesful for "winback" marketing efforts. This piece of code is installed on every user who deletes the Uber app, as Uber has no way of knowing if the iPhone will change hands at the time that the Uber app is deleted. So everyone who deletes the app can be tracked and presumably sent winback communications.

***

It is debatable whether consumers care about being tracked 24/7, and I don't pretend to speak for everyone. Uber is not the only company that has developed software to follow their customers' every move. What is clear though is that the engineers who write the code to execute these concepts - at Uber or elsewhere - believe that the consumers do not want to be tracked 24/7. However, they choose to do it anyway.

We know they know we don't want to be tracked. That's because the tracking code is typically hidden from view, and there is a general disclosure buried in Privacy Notices or Terms and Conditions that everyone knows nobody reads. If no one cares about tracking, then it can be performed in the open. Similarly, the cover-up operation to hide the illicit code from Apple engineers reveals that the engineers knew they were breaking the rules, and they did it anyway.

The Times article contained another revelation. Uber buys data from a startup called Slice Intelligence, who resells data from Unroll.me. Unroll.me runs a free service that helps people get rid of the clutter of spam in their email boxes - you grant permission to the company to peek into your mailbox, and pull out the unsubscribe links from various email lists. Well, it turns out that this service is a front for corporate espionage. Once inside your mailbox, their code gathers data about your purchases, and sells the data to companies: in this case, Uber buys data from Slice about its competitor, Lyft.

Here is actually one of the unspoken secrets of this "big data" industry. Unroll.me is one of many, many apps that are designed to collect data about our daily lives while fronting to be something else. I am pretty sure that the various receipt scanning apps (for expense reporting) are doing the same. I was told that weather apps are location databases that track all their users everywhere.

Again, it appears that the founders, managers or engineers who work for these outfits assume that their customers do not want to be tracked in such a manner because all such operations are hidden from view, and any disclosure is usually buried inside legalese that almost no one ever reads.

***

Slice Intelligence is hiding behind the weasel word "anonymized," which it explains as "no customer names." Usually, the deniers say they do not provide any "personally identifiable information" (PII), which would include phone numbers, addresses, emails, customer ids, etc. If any of those items are attached, not providing names is meaningless.

It is highly likely that Slice customers want personally identifying data - that is the key to connecting the dots. Analysts want to match John Smith on dataset A to John Smith on dataset B. If one of these datasets is truly anonymized, it will be very challenging, if not impossible, to correlate the data.

The "cover" of anonymized data is archaic, and I am surprised it is still in use. It's been proven a number of times that analysts can recover people's identities easily, even in datasets that are stripped of PII. Say, you are a loyal Uber customer, and pretty much take Uber cars everywhere. If I am given all of your trip information, origins, destinations and time of travel, I can immediately figure out where you live and where you work. From there, I will likely be able to identify you. Then I can look at the other trips to build a profile of your preferences, by analyzing what stores you shop at, what restaurants you eat at, how much you tip, etc.

***

Data sleaze is the data about one's own customers that are obtained secretly by businesses, and then sold to the highest bidders, also in secret transactions. The production of data sleaze is frequently justified by giving services away for "free." However, running a business as a "free service" fronting a profitable espionage operation is a choice made by management teams, not an inevitability. Indeed, many businesses that have a proper revenue model also produce data sleaze.

 

 

Dispute over analysis of school quality and home prices shows social science is hard

Most of my friends with families fret over school quality when deciding where to buy their homes. It's well known that good school districts are also associated with expensive houses. A feedback cycle is at work here: home prices surge where there are good schools; only richer people can afford to buy such homes; wealth brings other advantages, and so the schools tend to have better students, which leads to higher test scores.

A few weeks ago, the New York Times Upshot team ran some visual analytics to illustrate the correlation between school quality and home prices (link). Good for them, they went beyond a simple bivariate analysis. They also added commute time to the mix. The primary concern appears to be comparing so-called central cities with their suburbs in their respective regions. Some of these suburbs have better schools but the commute times are on average higher. Is it worth it to move to such neighborhoods?

This article has attracted much controversy. Alexander Russo summarizes some of the complaints here. The key objections cited in his report are as follows:

1. using test scores as a proxy for school quality is simple-minded
2. no accounting for other important variables such as demographics, backgrounds and so on (i.e. no attempt to control for other variables)
3. not causal: no one believes that moving a randomly selected student from a "bad" district to a "good" district will cause his/her test scores to jump by say two grade levels
4. using grade equivalent units which no one can understand

There are also complaints about the visual design. I will have some comments in a separate post on the sister blog. (The post should be available soon after this post goes live.)

***

I think #2 and #3 above are very important caveats. Point #1 is definitely a concern but critics admit that there are no great proxies. I will get to Point #4 below.

For me, the biggest issue is something missing completely from the original analysis: what decision scientists call the "indifference curves". What underlies this analysis is the assumption that families are making trade-offs between commute times, school quality and home prices. An indifference curve makes explicit what these trade-offs are: for example, how many more dollars per square feet is someone willing to pay in order to send kids to a school that is one grade level above another school? These curves are generally not linear. The value of moving from 3.5 to 4 grade levels is not as high as that of going from 1.5 to 2 grade levels, even though both shifts are precisely half a grade.

Once these trade-offs are in focus, we see the relevance of complaint #2. In the New York area chart, shown above, almost all of the districts are in the lower right quadrant relative to NYC, which means that a lot of districts on average have higher test scores and lower home prices than NYC. So, if those two factors were the primary reasons for selecting where to live, then no one would rationally choose to live in NYC! The addition of the third factor indicates that commute time may be another key consideration. Is commute time able to explain all of the data? That's not something easy to see from this chart, especially when commute time is reduced to two levels (under/over 30 minutes) and encoded in the colors.

***

The test score data came from the Stanford Education Data Archive (link). After you give them your email address, you will find a note saying that one standard deviation is roughly three grade levels. This raises a different kind of issue! Are the differences plotted on the chart meaningful? Usually, meaningful differences are at least two standard deviations, which would be six or more grade levels.

"Grade level equivalent" is a complicated concept. The same test is administered to students of different grades, so for example, the 5th-grade test is given to students in 3rd, 5th and 7th grades. The average scores of 3rd, 5th and 7th graders are computed. These are then mapped back to the 5th-grade scale. Anyone in 5th grade who scored above the average 7th grader is said to be two grade levels "ahead".

The fact that one standard deviation equals three grade levels tells me that the score distributions of 5th graders and 7th graders taking the 5th grade test overlap substantially. It seems likely that the difference between the 5th and 7th grader is not statistically significant - which, if true, is a sad commentary on our education standards.

Back to the chart above. If my interpretation based on the note from the Stanford Eduaction Data Archive is correct, then almost all of the horizontal differences on the NYT charts are practically meaningless.

***

Now, click here to read my comments on the data visualization.

 

 

My pre-existing United boycott, and some musing on randomness and fairness

You probably already saw the video - if not, do yourself a favor, and search for "man forcibly removed from overbooked United flight."

Other than the video evidence, which is damning, we don't have many facts, other than assertions made by various parties, repeated endlessly on social media and mainline media. Some facts, such as the United CEO claiming the passenger was "belligerent," is an assault on the meaning of the word.

United's claim that four passengers were "randomly" selected to be kicked off the flight, in order to make room for the four United employees, remains unchallenged. What procedure was used to select these four passengers, and is it really "random"?

***

What does "random" mean? Mathematically, we can say that every passenger has an equal chance of being selected as one of the four. According to SeatGuru, the airplane on that route holds 70 passengers. So the chance of any specific passenger being selected is 4 out of 70 (roughly, 6 percent). This is about the same as seeing four heads in a row in four coin flips.

The seating map shows three classes of passengers, six in First Class, 16 in Economy Plus, and the rest in Economy. The first unknown is whether passengers in different cabins are treated equally or not. My hunch is that First Class passengers are excluded from this lottery. (If you know anything about this, feel free to leave a comment.) It seems to me these airlines micro-manage all decisions based on fare - for example, the fare basis is the reason why you frequently get only one or two seats to pick from when booking but when you board the plane, you discover it is half empty.

The second unknown is whether passengers within the Economy cabin were treated equally or not. As is well known, even if you sit in the same cabin, you have not paid the same fare. According to this website, the median "restricted coach" fare for this United itinerary (operated by Republic) is $219 with a range from $80 to $426. For "unrestricted coach" fare, the median is $135 ($70, $246). For "restricted first class," it lists a median fare of $102 ($67 to $469). Are lower-fare passengers more likely to get selected?

The third unknown is which passengers had connecting flights and which didn't. From an operational perspective, bumping a passenger who has a connection can create a second-order effect, as this passenger will now miss that connection. If the selection was truly random, the airlines would have ignored this factor.

The fourth unknown is how airlines handled group travellers. Given that groups of travellers are not identified by a special code (as far as I know), they can't perform random selection at the group level. Therefore, what happens might be as follows: they select randomly at the passenger level, meaning that every passenger has 4 out of 70 chance of being selected; however, if any of the randomly-selected passengers is part of a group, United personnel face one of three choices:

(a) keep the selection and break up the travelling group;
(b) disqualify the individual who is part of a group, and replace him or her with another random selection;
(c) if the travelling group is not more than four individuals, bump the whole group, which means one or more of the other randomly-selected passengers will be "relieved."

Going with choice (b) destroys the randomness because any passenger who travels in a group has zero chance of being selected. Going with choice (c) also nullifies the randomness because passengers who travel in a group have more chance of being selected - if any one group member is selected, the entire group is selected. The larger your group, the higher your chance of being bumped. Choice (a) maintains the facade of randomness but predictably, passengers in the group typically want to stay together.

For these reasons, I don't believe the selection process was truly "random." (Again, if you know something, say something.... in the comments section.)

***

What would United do - in defence of randomness - if the randomly-selected passenger were a TSA "sky marshal" or, for argument's sake, Rahm Emanuel? Would they have respected the random selection?

There is a longstanding controversy amongst statisticians who use random sampling. When you are drawing randomly just one panel of people, the specific panel may be skewed. If you know that the randomly-assembled panel is skewed, do you resample?

Sometimes, random selection is not appropriate. For example, and I have mentioned this before, it is not plausible - nor is it conscionable - that the NYC police is inspecting bags "at random" in the subway system, as the announcers keep insisting to passengers. Truly random selection means inspecting only people picked out by a randomizing device. If one uses judgement to decide which bags to inspect, the process can no longer be described as "random."

***

Randomness is often equated with fairness. Random is "fair" only in the average sense. But lots of events only happen once. The one instance may or may not be fair, even if randomly generated. Throwing a fair coin four times, as stated above, has a 6 percent chance of showing four consecutive heads. If you are giving out four prizes by random drawing to a room with equal numbers of men and women, and all four prizes went to the men, the process is "fair" but the outcome may be perceived to be "unfair."

The video has become a sensation in China because of the race of the passenger who was dragged along the floor like a sack of potatoes. Minorities are very sensitive to such events, for good reason.

In order to understand why minorities are quick to blame racism (or race-based bias), you have to understand two modes of thinking about uncertainty. (See also the related blog post here.)

Here is a stylized description of the "majority" perspective: You hear that the selection is random, therefore fair. You accept that claim, as you generally believe that you live in a fair and just society. You are appalled at how United treated the Asian man. You also believe that United would have bullied anyone in the same way, whether you are a massive UFC fighter, or a beautiful model straight out of the cover of Vogue. You see the incident as an aberration in a fair system.

But now flip that framing over. If you are a minority, you don't generally presume the system is fair. You know it's less than fair; you just don't know if it is a little biased, or a lot. To answer that question, you look for data. The data include incidents like this assault on the defenceless Asian man. In a Bayesian way, you update your belief system after each new incident; you think the system is a bit more unfair than you previously thought.

This line of argument is covered in this old post about inverting probability. Assume we all observe that in four coin tosses, there are four heads. The mathematician will rely on the assumption that the coin is fair, and argue that seeing four heads is highly unusual, and occurs only 6 percent of the time. The statistician will judge whether the observed outcome of four heads is sufficient evidence to show that the coin isn't fair to begin with. In other words, the direction of thinking is inverted.

***

tl;dr

After the appalling incident in which an elderly, Asian passenger was forcibly dragged out of an aircraft like a sad sack of potatoes, the United management hid behind, among other things, a claim of "randomness." That claim has gone unchallenged. It is highly unlikely that the selection process was truly random, in the sense that regular people understand the word. In fact, there are many valid reasons for the selection not to be random. Statisticians have long figured out that "random" is not the elixir that solves all problems. The claim of randomness is also frequently used as a shield against allegations of bias. This incident will be seen as an aberration to those who presume we live in a fair and just society, and it will be seen, by minorities, as incremental evidence showing that the system is less than fair.

Also, to explain the title of the post: I have been boycotting United long before this appalling video. There was a time when I traveled enough to have "Premier Executive" status on United, but in recent years, I've woken up and said NO to further belittling and disgraceful treatment by sundry United personnel, including "customer service" people. #BoycottUnited

P.S. Don't let the cheap fares tempt you - it is not worth it.

P.P.S. According to the link provided by Dinesh in the Comments section, the "random selection" is a complete lie. United policy does not mandate anything like a random selection. This is the relevant text:

1. Passengers who are Qualified Individuals with Disabilities, unaccompanied minors under the age of 18 years, or minors between the ages of 5 to 15 years who use the unaccompanied minor service, will be the last to be involuntarily denied boarding if it is determined by UA that such denial would constitute a hardship.
2. The priority of all other confirmed passengers may be determined based on a passenger’s fare class, itinerary, status of frequent flyer program membership, and the time in which the passenger presents him/herself for check-in without advanced seat assignment.

The second bullet point provides a huge amount of room for discrimination.

What is Mr. Pruitt saying?

In his latest provocation, the EPA Chief, Scott Pruitt, aborted an on-going process by his agency to ban a widely-used but potentially harmful pesticide known as chlorpyrifos (link to New York Times article).

In my previous blog on his climate-change statement, I pointed out that people who attack data-driven conclusions for its "imprecision" will ignore any uncertainty if they want something to happen:

However, when it comes to such decisions [i.e. risky business investment decisions], the same decision-makers who fear the scourge of imprecision suddenly re-make themselves as "betting men."

Little did I know that it took just a week for Mr. Pruitt to prove me right.

The Times excerpted a part of the pesticide statement by Mr. Pruitt, as follows:

We need to provide regulatory certainty to the thousands of American farms that rely on chlorpyrifos, while still protecting human health and the environment... By reversing the previous administration’s steps to ban one of the most widely used pesticides in the world, we are returning to using sound science in decision-making — rather than predetermined results.

What is Mr. Pruitt saying?

Since he is not ordering more studies to dissipate the "imprecision," Mr. Pruitt knows something we don't. He has access to this "sound science" which proves that chlorpyrifos does not harm humans - with 100% precision. Obviously, he knows more than the scientists at his agency, who are concerned about its potential harm.

In addition, the disagreement between his agency's science and the "sound science" of unknown origin does not concern him at all. Didn't he just say that he does not believe in climate-change science because "there’s tremendous disagreement about the degree of impact"?

***

Mr. Pruitt claims he is providing "regulatory certainty." That phrase sounds more complicated than it is. When an agency makes a regulation, it is providing regulatory certainty, isn't it? Whether the regulation is to ban or not to ban a substance, it is still providing regulatory certainty.