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

Is A/B testing that scary?

Reader AR pointed me to this Fast Company article that examines the ethics of A/B testing.

The only way to comprehend this point of view is to think of A/B testing not as a scientific experiment but as a decision-making process that involves running an experiment. The researchers are unhappy that A/B tests could lend support to decisions that have undesirable impact on society.

Two such examples are described:

• Two images are tested for a job ad. During the test, site visitors were shown one of the two images, selected at random. The winner of the test is an image that disproportionately drives male applicants.
• Separate pricing tests are run in different zip codes. The "winning" prices at the conclusion of these tests are different for different zip codes. Because racial profiles differ by zip code, prices are in effect different for different races. Therefore, the test result leads to race-based discrimination.

There are two important questions to discuss here. First, what is the alternative to A/B testing? Is that method of decision-making better? Second, is the harm produced by the experiment itself, or by the decision made as a result?

 

Alternatives to A/B Testing

Consider the image test described above. Presumably, the test is run because someone believes that one of those two images might perform better at driving applicants. At most companies, a test sees the light of day after teams of people debate and prioritize testing ideas. If the test including a sexist image is run, then the team in charge of testing has approved it for some (possibly bad) reason.

If they didn't have A/B testing, how would they have decided which image to run? And if the image is not explicitly sexist - in other words, if the analyst had to analyze the data to learn that one image drove more male applicants - how would that insight be surfaced without running the other image? The alternative decision process may be even worse.

It is certainly true that automated A/B testing is risky - because no human beings are involved in turning test results into actions. The absence of humans is usually touted as a benefit by vendors of such testing tools. In this example, a human analyst reports on the test result, and includes the analysis by gender showing that while total applications increased, the winning image disproportionately attracted male applicants. The decision-makers can and should decide not to adopt the winning image based on that analysis. The A/B test revealed the bias but did not cause it.

Even without the gender issue, such analysis and discussion of results is necessary. For example, ad clicks can be generated by placing ads near scroll bars to stimulate accidental clicking. Human analysts can report that clicks increased but only through accidental clicking. The decision-makers can and should decide not to implement the winning design.

From where does the harm come?

The other example is more far-fetched. I am reverse-engineering the pricing test as described. Given that the test led to different prices for different zip codes, they would be running separate A/B tests stratified by zip code. Given the law of supply and demand, it might be the case that the winning price would be lower in poorer zip codes and higher in richer zip codes. This definitely results in price discrimination by zip code. If the design team did not want price discrimination by zip code, then such a test design would not have been approved so the test itself isn't creating harm.

Further, race-based price discrimination is accused because zip codes are correlated with race. Almost all variables are correlated with race. Age is correlated with race, so are income, education, what websites one visits, etc. So this standard leads to a banning of all segmentation and targeting policies. The only possible pricing policy would be one price for all.

***

In short, human supervision of A/B testing from design to interpretation is definitely needed. A/B tests provide a wealth of data to support decision-making. The biases highlighted by the Fast Company article are merely revealed by the testing - they are not caused by it.

 

Regulating data sharing is heating up

With the U.K. report on Facebook, and the stern language within it, the train on regulating data sharing may finally reach the station this year. The FTC is also likely to impose a stiff fine on Facebook for violating a consent decree.

So let's learn more about this data sharing business. If you prefer a video, the gist of this post can be heard here.

***

First, let's talk about data flows and the "cloud". Data are stored in computers that are called servers. In the cloud computing model, these servers are owned - not by the companies that collect the data - but by large tech companies like Amazon, Google, Microsoft, etc. who are responsible for managing the servers. These servers are geographically dispersed and so when data enter the cloud, they get replicated and spread to many servers. The technical benefit of such replication is recoverability of the data (allowing the use of cheaper, less reliable computers) but now, the data become much harder to delete.

Data become more telling if one combines different datasets measuring different aspects of our lives. For example, an auto insurer may have data on past claims and that data help predict your future claims. But if the auto insurer is able to get data from say an automaker about your car, e.g. how fast you drive, where you drive, etc., that data combined with past claims improve the predictive power.

Thus, a data-sharing industry has been created. Companies make agreements to share data with one another. This becomes much easier in the "cloud" as those servers are already connected to one another. These agreements may include explicit payments but even if they don't, both sides must be benefiting commercially from the arrangement, or else they would not exist.

So when company A shares data with company B, the data flow from A servers to B servers. B may also use a cloud, which then means the data would be replicated yet again, and dispersed geographically onto yet another set of servers. 

And company B may also share data with company C, etc., etc.

***

An inexplicable part of the consent decree between Facebook and the FTC is the requirement that Facebook monitor what happened to the data after they are shared with third parties. I just can't figure out how that is possible. It isn't even possible within Facebook: if a user demands that his/her be deleted, it will be very hard to ensure that all copies of the data are deleted from every server, including data that might have landed in an analyst's computer. In fact, most analysts probably don't know how many replicates of data elements are being created during the analysis, and where those replicates exist!

***

The next question of general interest is all the different ways in which tech companies collect people's data without people realizing what's happening. In the video, I look at contact lists, personality tests, 2-factor authentication schemes, IOT devices, etc. in their roles as data collectors. 

This is the reason why the video is called "Did you betray your friend today?"

How scary are the anti-vaxxers?

I don’t agree with Daniel’s conclusions in his article in Slate about the measles “crisis” but he did his research and there are lots to chew on. You don’t have to agree with him to find this article thought-provoking.

There is one paragraph which everyone should read. It’s a celebration of science, and how it saved lives. (Daniel used this story for a different purpose: he argued that we never “eradicated” measles, and therefore, the anti-vaxxers could never have reversed some mythical victory.)

During the most recent, major wave of measles infection in the U.S., between 1989 and 1991, close to 56,000 people fell ill and more than 100 people died...The 1989–91 epidemic was large enough and deadly enough to cast light on two pressing problems: First, that a single vaccine dose was not sufficient to protect children, and second, that black and Latino children, especially those living in urban areas, were less likely to be vaccinated, and thus more vulnerable to the disease.

Efforts were made to address both issues in the years that followed. A second measles shot was recommended for all children, while the federal government ramped up efforts to provide free vaccination to high-risk groups. The plan worked. By 1994, vaccine coverage for measles was closing in on 90 percent. The number of cases reported every year soon dropped from the thousands into the hundreds, and then into the tens. It was in response to this decline that experts from the CDC announced that measles had been “eliminated” from the U.S.

It's a great example of finding the drivers behind the data, and executing actions that successfully changed the numbers.

***

From the rest of the article, here are some useful tidbits:

• Vaccines work because of a phenomenon called “herd immunity,” which is a type of wisdom of the crowd. Diseases spread when people interact with each other. If both sides are vaccinated, then the risk of spreading is much, much lower. Thus, the higher the proportion of the vaccinated, the lower risk for everybody. The threshold desired by health authorities is 93 to 95 percent.
• In the past several decades, at a national level, the vaccination rate has stayed around 91 percent. So it is below the desired level but seems close enough not to cause alarm.
• Discredited research by a Dr. Wakefield ignited the anti-vaxxer movement. The publication of such research allows people to confirm their prior beliefs, and it often is hard to dislodge such beliefs, even when the research has been invalidated.
• In some localities, e.g. Somalis in Minnesota, the rate of vaccination has dropped drastically to below 50%. Those are isolated communities, and in aggregate, the level of vaccination has not changed.
• The number of measles cases while small has shown signs of increase. There have been triple-digit cases 7 out of the last 10 years, but only twice in the previous decade.
• The fatality rate of measles is extremely low, 11 deaths in 18 years, which is said to be similar to being killed by scorpions. Those who read Chapter 2 or 4 of Numbers Rule Your World (link) will recognize the need to think about the cost of errors. 

 

Roundup of AI startups

Business Insider profiles 12 AI startups that their panel of venture capitalists considers likely to succeed in 2019.

These startups fall into three categories:

A) Enterprise apps: Appzen (auditing), Atrium (legal), FortressIQ (processes), Guru (knowledge management), People.ai (sales)

B) Robots: Farmwise (tractor), 6 River Systems (warehouse), Shield.ai (drones in risky places)

C) Others: Superhuman (email app), SambaNova (chips), Transfix (marketplace for trucking)

***

Of these, I think group B (robots) is the most promising. The self-driving technology is particularly well suited in these settings (farms, warehouses) in which traffic control can be fully centralized. In some cases, the accuracy required is not too high, e.g. the tractor that differentiates weed from not weed needs only be moderately accurate.

The drone company - it's not clear where the AI is.

In Group A, I like Appzen, which uses AI to detect fraud in expense reports. It's clearly possible, and has a business model - although the model is about cost control rather than revenue growth.

Guru, People.ai, and FortressIQ are applications that have current solutions, and it's not easy to see from a brief description how they will be better. Knowledge management and process streamlining are problems that are around forever, and never get solved.

Atrium for legal documents is too vague a concept to judge.

Superhuman, the new email interface, can be nice but Gmail killed the market for email software by giving it away free. AI is likely just a small part of this offering. I must say I like the new Gmail predictive typing feature. On the other hand, Gmail's predictive classification of emails to tabs continues to underwhelm after years of training.

SambaNova has a great team, but no details on what they are really doing. Transfix, the marketplace connecting truckers and loads, sounds promising - if only because this problem has been studied for a long time.

At the end of the year, we can come back and see how these companies performed.

When the ATM fails to read my checks

Happy Lunar New Year! And greetings to Orlando people who are coming to my dataviz seminar this morning.

***

What’s going on with digit recognition, one of the signature applications of machine learning?

Before self-driving cars, before image recognition, before machine translation, there was digit recognition: computers are trained to read and recognize hand-written numbers. This problem shares several of the key components of problems tailor-made for machine learning methods:

1. The correct answer is unambiguous for each item (i.e. image of a digit). The author of the digit has a particular number in mind.
2. The range of possible answers across all items is finite. In a decimal system, each image can only be one of 0, 1, 2, ... , 9.
3. The end-user only cares about how accurately the digit can be predicted. Causality is not of interest here.
4. A massive dataset of labeled images, i.e. images that have been correctly recognized, used to train computers is easily obtained.
5. Live application generates more data, which feeds back into the system in a positively-reinforcing manner.

This digit-recognition technology is widely deployed. Check deposit at the ATM machine is one obvious example. In 2016, about 16 billion checks are deposited in the U.S. (source). So what’s wrong with the current state of art?

This snapshot I took at an ATM illustrates the problem:

Recently, I noticed that the ATM has refused to recognize the digits on several checks, asking me to enter the amounts manually instead.

From this evidence, I infer the following:

1. Still after these years, the error rate is higher than these banks could absorb. Assuming 10 billion checks read each year at ATMs, even a 0.01% error rate amounts to 1 million errors per year, or about 2,800 errors per day!
2. Banks would rather err on the side of caution – when in doubt, ask users to enter the amount. This behavior implies humans make fewer errors than machines, even after including mischief as a source of human error.
3. What would a teller do if s/he can't make out the scripted digits? The human would look at the handwritten words "six thousand," solving the problem. Apparently, the ATM does not have handwriting recognition technology, or perhaps its accuracy rate is not high enough. It's a harder problem, though of a similar nature.

***

Why are the banks risk averse? As a victim of one of these errors, I think I understand. Last year, I spent four or six weeks chasing after $20. In this case, the machine read the 2 as a 0. I didn't catch the mistake while at the ATM, but later noticed it on the bank statement.

I learned that convenience comes at a price. The bank's process to verify the amount and correct the mistake is convoluted. It's like missing that exit on the highway, and you now have to go five miles before the next exit. It's a pain for the bank as well as for the client.