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

What's wrong with the Dropbox research?

In a mailing list I subscribe to, some users were not happy about the academic research using Dropbox data collected on academics, as written up by the co-authors in Harvard Business Review (link).

In a nutshell, the researchers obtained "anonymized" "project-folder-related" data from Dropbox on university-affiliated accounts, and did some simplistic bivariate correlations, and proceeded to draw several conclusions about "best practices" on "successful team collaborations." This type of research is very common in this "Big Data" age, and I have already written extensively about its many challenges.

OCCAM Data

This Dropbox dataset has all five descriptors of the OCCAM characteristics of "BIg Data". It is Observational, seemingly Complete, with no Controls, Adapted from its original use, and Merged (with data from Web of Science). These characteristics cause many problems with the analysis, which I describe below. For more on OCCAM data, check out this post, and my other posts on OCCAM data.

Found Data

Implicit in their analysis - and most other uses of "found data" - is the assumption of complete data. The authors believe that because their data consist of tens of thousands of researchers and 500,000 projects, they must have all the data. In this case, the authors knew that there are other platforms out there but they waved away the inconvenience. This implies they believe they have all of the "informative" data.

The authors also assumed that (a) all relevant collaborative research involves putting all relevant files on one of the major online platforms (e.g. nothing on Slack or emails) and (b) all project collaborators use one and only one platform. Further, they assume that everyone has highly organized and structured folder directories within Dropbox for which an external person who knows nothing about the projects, or a machine, can infer its contents. These problems arise because the researchers did not start with a research question and design the data collection. They chose to adapt "found data" to their own objectives.

Ecological Fallacy, and Story Time

A typical conclusion is "People at higher-performing universities seemed to share work more equally, based on the frequency of instances that collaborators accessed project folders." Top universities are ranked by the aggregate performance of all teams (that use Dropbox, and are identified correctly). It does not follow that every team at a top university is a top team.

This is an instance of "story time." A piece of data is offered about something related, then while the reader is dozing off, it is linked to a conclusion that is not directly supported by the data. That conclusion is elaborated by a lot of words hypothesizing why it must be true. In this case, they say "It’s likely that more frequent collaborations led to positive spillover of information, insights, and team dynamics from one project to another." But they provide no evidence at all for this last statement. That's just a story.

Proxy Unmasking

Here's another one of the conclusions: "People at higher-performing universities seemed to share work more equally, based on the frequency of instances that collaborators accessed project folders." They make a conclusion about work allocation between different collaborators but what they actually measured was the relative frequency at which collaborators accessed the project folders in Dropbox. That's a proxy measure, and it's convenient based on "found data", but not a good proxy.

Xyopia

It does not appear that a multiple regression model was run. The presentation walks through apparently a series of bivariate analyses. The word control does not appear in the entire article. So this work suffers from xyopia - in each analysis, the one explanatory variable being analyzed is presumed to be the chief and only variable that influences the outcome.

Causation Creep, and More Story Time

The authors made no attempt to establish causality at all. They just interpreted every correlation as causal. So every conclusion is "story time". They print one analysis of the data, then they draw a causal conclusion that one would believe only when half-asleep.

***

People are also upset about data privacy.

1. It does not appear that the academic users understood that using Dropbox means they are part of research studies.
2. People don't believe the data are truly anonymized. It's pretty clear that the anonymization can be easily reversed. Just take the HBR article for example. If they removed the names of the authors but retained information about the authors as: one junior faculty at Northwestern Univ. Business School, one senior faculty at Northwestern Univ. Business School, and one employee of Dropbox - I don't think you can find another article that fits those criteria. So is that anonoymous?
3. It's unclear if how they "anonymized folders" or analyzed them. There are folders with highly descriptive names, there are folders with partially descriptive names that only the collaborators may be able to decipher, and there are folders with names that do not identify the project (e.g. old_work). If they converted all folder names to alphanumeric strings, then all information about the contents of the folders is lost. If they don't convert those names, then there are clearly privacy concerns.

It's clear that some kind of IRB review is necessary to approve Big Data research projects to make sure privacy is protected.

Know your data 23: Hands on your emails

The media have finally started to write some really nice reports on data sludge. I like this Wall Street Journal article, opening the black box of the science of secretly reading your emails.

If you use a smartphone, it is very likely that you have agreed to some app's terms and conditions, which allow them to download your emails en masse from one or more of your favorite cloud email providers, such as Gmail, Yahoo! Mail, Outlook, etc.

There was already an infamous example of this that came to light last year. Unroll.me is a service that helps you unsubscribe to unwanted mailing lists. When you set up this service, it requests access to your emails. That is the way they find out which services you can be unsubscribed from. It turns out that Unroll.me isn't really about helping you reduce email clutter - in fact, its main business is mining your inbox for shopping receipts, which can be sold to businesses which - you guessed it - want to sell you more stuff, which - you guessed it - probably means you'll receive more spam, net net. Oops. That was the sound when the company's management learned their little data sludge scheme went public.

For the Unroll.me story, I'm linking to this commentary in Venture Beat by someone who audaciously slammed Unroll.me management for audaciously claiming that their data sludge scheme was par for the course in the tech industry. This guy actually screamed: "The analogy between Unroll Me and Google or Facebook is audacious. Not to say haughty." He went on to claim that Google and Facebook keep all their data in-house. That was false in 2017, and looked even worse in light of recent revelations about data practices at those two companies.

Google, for example, has allowed, and recently expanded access by third-party developers to Gmail emails, according to the Wall Street Journal. Just like Facebook, Google has no control over how these third parties use the data. It has some language requiring these developers to agree to certain standards but those are unenforceable, and not enforced.

The WSJ article includes quotes from various participants in this data sludge industry that are false, intentionally or not.

First, they repeatedly claim they don't "read" our emails. Let's do a thought experiment here. I want to know if you are a racist. Unbeknownst to you, I got my hands on all the emails in your Gmail account stretching back 10 years. I write a computer program to look for various keywords like the N word. The program tabulates for me how many times you used each word, which days of the week you tend to say such words, which people you use those words with, the number of variations of each such word you have in your vocabulary, how many of your friends partake in such conversations and how many times they use racist terms, etc. Based on this report, I conclude that you are a racist. I may even conclude that certain friends of yours are also racist. According to various interviewees in the WSJ article, I drew that conclusion without "reading your emails."

(Lest you think the example is far-fetched, we recently heard that Facebook had tagged thousands of users with the label "treason," a segment which can be purchased by advertisers - or anyone willing to pay for this data.)

Second, the companies interviewed for the article e.g. Return Path basically claim that they have only had human beings read emails once or twice. That is simply a lie. You can't build any kind of predictive model without getting intimate with the data. Further, to understand how these models work, you have to review actual cases. Finally, when something unexpected happens, you have to look at the email contents to understand why.

These technologies have some possible benefits. If such benefits outweight the potential harm, then consumers would gladly adopt them. The data industry should be much more transparent. This ensures that the developers maximize benefits while reducing the levels of harm. 

***

We've been tracking data sludge for years. For more, read this thread.

Academic trickery, annotated

In the last post, I linked to the article by Philadelphia Inquirer, disclosing how admissions staff at Temple University gamed rankings. They mentioned specific techniques. Here is an annotated version, with my comments:

Inflating GMAT Scores

The most interesting technique is converting GRE scores to GMAT scores before computing and reporting the average GMAT score. Presumably, those taking GREs are better students because they are applying to MS and PhD programs which require GRE and do not allow GMAT as a substitute. So, by such conversion, the average GMAT score is inflated, according to the rules which require including only those applicants with GMAT scores.

Is converting GRE to GMAT a bad thing? Not necessarily! The average GMAT score only represents the part of the applicant pool that submitted GMAT scores. As discussed above, this group is likely to be less academically brilliant than the other group who submitted GRE scores. (Chances are the average GRE score is separately reported.)

What is making trouble is the proportion of applicants submitting GRE versus GMAT scores. If a school has lots of applicants who are applying to MS/PhD programs and they submit GRE scores instead, then the average GMAT score will likely to be dragged down.

I actually think having a standard conversion formula between the two tests is great. All schools can then be evaluated on one average test score that takes into account the mixture of GREs and GMATs. The question is: is there a standard conversion formula?

Yes, the ETS provides one. This formula is based on the subset of people who have taken both tests. You can then use one score to predict the other score. Here is a PDF that explains the methodology. (By using this formula, we make the assumption that this subset of test-takers can be generalized to the subset of test-takers who took the GRE, did not take the GMAT, and are applying to business schools.)

Under-reporting student debt

The reported average student debt was diluted by including students without debt. The rating body has requested that the metric be computed only for students with debt. Mixing in zeroes brings down the average. 

There are two different averages to be considered: the average debt held by students who have debt; and the average debt for all students. If we want to evaluate the school's financial aid policies, the average debt for all students gives a better answer. If you are a prospective student who will be taking out a student loan, the average debt held by students with debt is a better measure of the amount of loans required.

The link between the two metrics is the proportion of students who have debt. If the school's policy is to "spread the wealth", then the average debt load will be lower but a higher proportion of students will have debt.

Rounding up GPAs

It appeared that they crudely rounded up the average GPA. The example was rounding 3.22 to 3.30. The 0.08 increase looks innocent but applied to the average, this means adding 0.08 to everyone's GPA (to be more accurate, for each student whose GPA is 3.92 or higher, someone else's GPA got inflated by more than 0.08).

Under-reporting the number of admission offers

Selectivity is the number of offers divided by the number of applicants. Inflating the number of applicants or deflating the number of offers increases selectivity rate. According to the investigation, they blatantly lied by under-counting the number of offers. In the previous post, I described a number of techniques that are more subtle - you generate more applicants but of the kind that you are unlikely to give out offers to.

For even more tricks, read Chapter 1 of my book Numbersense.

 

Real life imitates book

In Chapter 1 of Numbersense, I describe a bag of tricks a statistician can play to "game" school rankings. The obsession about these rankings is unhealthy - but the reality is that people cling to imperfect measures even when they are known to be imperfect - it's the something-is-better-than-nothing school of thought. Because parents and prospective students use these rankings, school administrators have to pay attention to them.

In the book, I looked at the impact of the Common Application form in college admissions in which a student fills out one form, and sends it to a set of schools. I also discussed the effort to reach out to under-privileged sections of society by elite universities. Both of these efforts have the effect of increasing the number of applicants much more than they are able to raise the number of accepted students, thus pushing up the selectivity rate.

An item just appeared in Princeton Alumni Weekly which illustrates the math well.

Until this year, Princeton has not allowed transfer students for a long time.

The admission rate amongst transfer applicants is 0.9 percent (13 out of 1,429). The admission rate for all applicants is reported to be 5.5 percent (see here). Therefore, if we mix transfer and regular applicants, the blended admission rate is pulled lower. In this case, it becomes 5.3 percent.

The math part is not controversial. If any school adds applicants, and especially if it recruits applicants who have a lower-than-average chance of getting admitted, such as under-privileged communities, or transfers, such policies directly lead to a numerical improvement in selectivity.

However, I am not criticizing the transfer policy for the following reasons:

• The math does not tell us about intention. Rising the selectivity rate is either an intended or an unintended consequence of the policy. I'm not making a judgment.
• Most other universities already have a transfer applicant policy so their blended admission rates already are deflated.
• Rating or evaluation bodies probably require excluding transfers from the official selectivity rate (although I'm not sure about this).
• If the proportion of transfers is small (in Princeton's case, the first year transfer applicant pool is about 4 percent of total applicants), the transfers may have a negligible impact on the average. (On the other hand, for rating purpose, even a small percent shift matters.)

***

In the meantime, I'm watching the scandal around Temple's Business School. The dean there seemed to have acted like the Devious Dean in Chapter 1 of Numbersense, finding various ways to falsify the data sent to U.S. News. Here is the Inquirer's take on the developing scandal.

 

 

 

A puzzle, and a cause for concern

Last week, I referenced the Wall Street Journal article titled "You Graduated Cum Laude, So Did Everyone Else," published on July 2nd.

When I wrote up the blog post, it was July 4th, two days later.

I couldn't find the article on Google News. Even when I input the entire title of the article, Google News still came up empty. Here is the screen shot:

It found irrelevant articles from WSJ, some from months ago but it did not retrieve the most recent relevant article - even when I wrote the entire title of the article!

I also tried other queries like "wsj melissa korn" (author of the article) but obstinately,this one article just refuses to come to the surface.

The problem is specific to the algorithm behind Google News. When I tried the exact same query on Google's search engine, the first result coming back is the correct one:

 

Next, I experimented on Bing. Both Bing News and Bing Search return the expected result.

***

The puzzle: what is confusing Google News? how did the algorithm miss the most obvious result?

The cause for concern: the algorithm behind Google News (and other news sites) controls what we see and what we don't see. We have essentially zero information about why the algorithm selects what specific articles for us.

Comment below if you have thoughts about this.

 

Nothing matters but straight As

Fifteen years ago, Princeton was famous (infamous, depending on your views) for being a leader among elite U.S. universities in the fight against inflating grades, imposing a quota on how many As professors are allowed to give out in any course, a policy described as “grade deflation.” The policy was scrapped in 2014, when the administration bowed to pressure from parents and lack of reciprocation from peer institutions.

It’s come to this. Fifteen years later, Princeton appears in the lead sentence of a Wall Street Journal article on the meaninglessness of GPAs and cum laude honors in the world of education in this millennium.

I have written a few posts about Princeton's losing fight against grade inflation (here, here, here). In 2014, I grudgingly understood the administration’s decision to abandon grade deflation as a bow to pragmatism, although I am not a fan of the policy reversal. So I won't repeat those arguments here.

But what about the journalism!

This is the entire first paragraph from WSJ:

Nearly half of students who graduated from Lehigh University, Princeton University and the University of Southern California this year did so with cum laude, magna cum laude or summa cum laude honors, or their equivalents. At Harvard and Johns Hopkins, more got the designations than didn’t.

So the paper decided to lead with Lehigh, Princeton and USC, where fewer than 50% of the graduates received honors, instead of Harvard and Johns Hopkins, where about 60% of graduates received honors. What motivates this choice?

The mention of Harvard merits further investigation because some years ago, it came out that almost everyone is an honors graduate, with only ~10% not given cum laude honors. See my previous post "Harvard gives new meaning to meritocracy." Did that change? Apparently so. According to the article, Harvard capped the honors class at 60%, starting in 2005. This Crimson article reported that the proportion of A grades did not decrease after the honors cap; and the college all but abandoned efforts to fight grade inflation.

A key sentence in the WSJ article is:

"A 4.0 does signal something significant, that that student is good," said Stuart Rojstaczer, a former professor at Duke University who studies grade inflation. "A 3.7, however, doesn’t. That’s just a run-of-the-mill student at any of these schools."

Note that if the school does not grant A+s, a GPA of 4.0 means straight As. He's saying anything but straight As is meaningless.

***

When we search for meaning for something like GPA 3.7, we are grasping for reference points. 3.7 at one school is different from 3.7 at another school; within a school, 3.7 in engineering is different from 3.7 in the humanities; 3.7 by one professor is different from 3.7 by another.

Understanding the distribution of grades at various levels is key to interpreting GPAs. However, few college transcripts are annotated with data about grade distributions.

Grade curving aids interpretation by re-scaling the data to well-defined reference points.

Moviepass stock plunges

Last time I mentioned Moviepass in May, the stock of its owner (NASDAQ: HMNY) was over $2. Today, it is nearing 20 cents! Lost 9/10 of its value in two months.

This latest plunge is related to a "hail mary" attempt to raise cash to fund its giveaway business model. It currently claims 3 million users and is losing $40 million ... every month!

Other news: the movie chains launched their own rival service. That's a smart move - kill off the value destroyer while it is still weak.

In the meantime, what does Moviepass say they are doing to keep alive?

One idea is "surge pricing": they are to impose surcharges for popular movies at popular times. And they have floated various other ideas to "upsell", generating more revenues from existing customers. It doesn't seem to realize that many of its customers are not the average movie-goer but the nickel-and-diming customers who smell a good deal when they see one, and will walk away if the deal isn't there.

The fundamental problem of the Moviepass business model has never changed - it loses money on all new customers, because it is paying full price for the tickets. Instead of fixing the problem on the new customers, it is trying to get as many customers in the "store" and then squeeze more money out of them while they are inside.

According to this Bloomberg article, the CEO explains that the strategy is "to sign up more subscribers who don’t go to the movies quite so often, making them more profitable." So he is betting on innumerate customers - people who don't understand that if they are paying $9.95 a month, they have to see at least one movie a month to make the subscription worth it.

Now, imagine you subscribe because you smell a good deal. The first movie you see is paid for by the monthly subscription. The next movie you see has to be paid for by another subscriber who has an idle month ("subscribers who don't go to the movies quite so often"). The third movie you see has to be paid for by another subscriber's idle month, etc. etc. If you see 10 movies a month, then Moviepass has to find 9 idle months to break even on your one subscription. And that is one month. The next month, if you see 10 movies, it has to find another 9 idle months to pay for them. (I'm using 10 movies not the maximum of 30 movies a month since I don't think there are 30 movies a month that someone can pay me to watch.)

Good luck to those who's still holding the stock.

A Report from an AI Day in Hoboken

Two weeks ago, I attended an AI Day organized by SAIL at Stevens Institute of Technology in Hoboken, and attended talks by AI staff from Google, IBM, Verizon, SAP and Verisk Analytics.

Here are some notes that I want to share.

1. AI is hot

All these companies are putting significant resources into AI. The people working in this field are fanatically excited, and hungry for a future run by software. They presume the world will be changed completely, and for the better, as AI infiltrates practically all aspects of human life.

2. AI is deep learning

AI has become synonymous with deep learning. Speakers use those terms interchangeably throughout the day. Deep learning is a direct descendent of neural networks, research that can be traced back to at least the 1960s, and thus it represents an evolution rather than a revolution. It’s hard to reconcile how this limited technology would produce the grandly stated ambitions.

Two important articles about the risk of betting everything on deep learning have appeared recently, one by Michael Jordan, the Berkeley data science professor, and one by the Allen Institute by way of New York TImes.

3. Different definitions of AI

AI is most effective and useful in automating repetitive, tedious tasks. For example, the police in many jurisdictions are driving around taking photos of license plates to compile a database of drivers’ locations. Instead of having humans read off license plate numbers one by one, they deploy an AI tool to efficiently read those numbers.

A telling moment at the conference was, after one speaker presented an example of using AI to take over tedious tasks previously done by humans, two people in the audience proclaimed that what he presented is not AI.

Later, the speaker from IBM declared all of today’s AI “narrow AI”, and indicated that IBM is focused on making “broad AI”. Broad AI is defined as technologies that are generalized enough to perform different skills, not just one skill. For example, today’s deep learner is trained to do one thing well, like determining if an image contains a dog or not.

What is AI? It seems like everyone has an opinion.

4. We know bias, and we will eliminate it

Another moment of the day is when one speaker turned to the conference organizer and said “It’s become obvious that we need to have a bias seminar. Have a single day focused on talking about bias in AI.” That was his reaction to yet another question from the audience about “how to eliminate bias from AI”.

As a statistician, I was curious to hear of the earnest belief that bias can be eliminated from AI. Food for thought: let’s say an algorithm is found to use race as a predictor and therefore it is racially biased. On discovering this bias, you remove the race data from the equation. But if you look at the differential impact on racial groups, it will still exhibit bias. That’s because most useful variables – like income, education, occupation, religion, what you do, who you know – are correlated with race.

5. It’s too good it’s bad

When it comes to ethics and potential harms arising from AI, the prevailing mood seems to be that bad is the consequence of AI being “too good.” There isn’t talk of curtailing the technology – the hope is that whatever bad is eliminated (see point #4).

This reminds me of the advice we’re given when answering the interview question: what is your greatest weakness? We’re told that a good answer is something along the lines of “I work so hard I forget to sleep.” We’re told that this answer is great because it does not answer the question, and in fact, switches the topic to something positive.

6. Real challenges

Between the lofty visions presented by various speakers, we get a few glimpses of what is actually being worked on, and the real challenges in deploying AI.

One speaker summarized the three key obstacles in deploying AI in the real world as “not robust, not explainable, and not trusted.”

The speaker from Verizon described how they are training an AI system – a kind of chatbot – that handles Q&A with customers. The complexity of such a system is clear, consisting of some advanced AI technologies running alongside rules-based heuristics.

Someone in the audience asked how this chatbot learns from its mistakes. Actually, before learning from, it has to learn of. Imagine you’re using this automated Q&A tool. You asked a question and it returns an irrelevant answer. How does the chatbot know it made a mistake?

The naive answer is: just display a survey question at the bottom of the response, asking the user whether the response is useful or not. This is a standard feature of almost all such systems that I’ve seen. What’s the reality? Few people complete this survey (sparse, biased data), and some mis-use the survey to express their general unhappiness (noisy data).

Instead, the team at Verizon randomly selects 100 answers a day that they examine manually for accuracy. This suggests human input is still needed to calibrate these systems on a continuous basis.

7. Where is the user in this?

More than once, I wondered what is the role of the potential user in the development of these AI technologies. In the talks that I attended, I barely heard mention of user testing, gathering requirements from users, collecting feedback from users, measuring user satisfaction, etc. The prevailing philosophy is if you build them (well), they will come.

AI is obviously an important trend to keep track of. Comment below if you have reactions to what’s happening.