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

How a fraud detection algorithm conspired to ruin my recent trip

I have been traveling quite a bit lately, and last week, I went to Rome for a few days, and spent time at the KDIR conference. Rome is one of my favorite destinations and apart from the architecture and museums, and the restaurants, I also enjoy shopping there.

To my dismay, a gray cloud followed me around this entire trip - in the form of a misfiring fraud detection algorithm. On foreign trips, I always prefer to spend cash via my ATM card to avoid those ridiculous credit card surcharges. And I have used my Schwab card without any issues for many trips.

This time, however, my card was blocked after one withdrawal. This means I had to call back to the States to unblock it. Then, the next day, it happened again. And this continued every day until the last day of the short four-day trip.

Needless to say, I was getting more and more irate as the trip progressed. It makes me wonder about fraud detection algorithms. Are they any good? If they are tuned to be very risk-averse, then you can always prove to your boss that you have prevented a lot of fraud. The flip side is you have always caused a lot of hassle for your good customers. In technical terms, each time my card was blocked, the algorithm committed a false-positive error.

The service reps had not a clue how these algorithms work. On the first day, I was told that it blocked me because I didn't give them warning that I'd be traveling. That made sense until it blocked me again the next day... after I told the rep the exact days I'd be in Rome. The next rep explained that I was getting blocked because Rome is a high-fraud zone, and I was using certain ATM machines. That sounds reasonable, except if those were the reasons, then I might as well throw the card away. The experience got me thinking about the challenges of making a good fraud detection algorithm.

Clearly, when I am traveling, my habits don't match what is in my customer history. I'm going to be engaging in a series of transactions that might look suspicious - like taking more cash out than usual, taking cash from places and machines that I have never used, taking cash out multiple times a day (because there is a per-transaction limit on most ATMs), taking cash out from machines all over town, etc. How can a computer figure out if those transactions are legitimate?

When the algorithm got it seriously wrong, it can be very annoying. One of those days, I had put money down on a suit. It was an hour away from the store's closing time. Because the problem could not be resolved in time, I had to go back the next day, which meant I had to cut other things out of the itinerary. If it happened on the last day of the trip, it would have been a lot of trouble. I racked up probably $100 of international roaming charges for all the calls I had to make to unblock the card repeatedly. There were several moments I had to stand on the street, the phone on one hand, the other hand operating the ATM, testing the machine, pulling out cash, etc. Those moments felt very ironic because the blocking of my card was supposed to make me feel secure.

As a statistician, I want to know the probability of falling victim to the kind of fraud Schwab's algorithm is trying to prevent, and the average cost of such fraud (bearing in mind that you can only take 250 euros per transaction). I suspect that the cost of the inconvenience, both tangible and intangible, may outweigh the potential benefit.

 

Count data are less useful than you think

A lot of Big Data analyses default to analyzing count data, e.g. number of searches of certain keywords, number of page views, number of clicks, number of complaints, etc. Doing so throws away much useful information, and frequently leads to bad analyses.

***

I was reminded of the limitation of count data when writing about the following chart, which I praised on my sister blog as a good example of infographics, a genre chock-full of deplorable things.

On the other blog, I explained why I prefer to hide the actual numbers, from a dataviz perspective.

There is also a statistical reason for not drawing undue attention to the counts.

These counts do not indicate the severity of the injuries: some may have knocked the player out of the game, others may have been much milder. Some injuries may be sustained by first-team players who spend a much longer time on the field than backups, thus raising their rate of injuries.

Another statistical consideration is heterogeneity. I'd like to see a small-multiples version of this chart, with the data split by position on the field. I think it will be quite telling which body parts are hurt more depending on one's role in the game. Similarly, splitting by age, body size, and other factors will yield interesting insights.

***

At about the same time, I was reading the July issue of Significance magazine (an RSS and ASA publication). Here is the link (not free).

In an article about assessing whether iceberg risk was particularly high in the year of the Titanic, the authors quantified the risk in terms of "number of icebergs crossing latitude 48 N each year". It'd seem worthwhile to ask whether there is also a relevant size distribution.

Then, in an article about "black box modeling" (i.e. data mining) by Max Kuhn and Kjell Johnson, they invoke an example of the FDA adverse event reporitng database, an example of "events data". Events data is everywhere these days, and the most popular analyses of such data revolve around counting the number of adverse events. The severity and type of events are frequently ignored.

 

P.S. In their otherwise gung-ho article, Kuhn and Johnson also point to one of the biggest challenges of OCCAM data. "If there is a systematic bias in a small data set, there will be a systematic bias in a larger data set, if the source is the same."If one is analyzing the FDA adverse events database, one must hope to apply the learning to people who don't yet have adverse events, but then such an analysis would be flawed since the database doesn't have any controls, i.e. people without these adverse reactions.

 

Pondering OCCAM data in medicine

The New York Times Magazine has a pretty good piece about the use of OCCAM data to solve medical questions, like diagnosis and drug selection. I'm happy that it paints a balanced picture of both the promise and the pitfalls.

Here are some thoughts in my head as I read this piece:

• Small samples coupled with small effects pose a design problem in traditional clinical trials. The subjects of the NYT article claim that OCCAM data can fill the void. If a treatment is highly effective, even small clinical trials will find the effect. So the underlying issue is less sample size than effect size.

• Counterfactual evidence is almost always absent from OCCAM data because of lack of controls (the first “C” in OCCAM). The lede in the story concerns a girl who was given an anti-clotting drug because a doctor suspected she had elevated risk of blood clotting, and the girl did not develop a clot. Statisticians are not impressed by such evidence, because we don’t know whether the drug was truly responsible for the outcome. (It's a correlation until proven guilty.) If the girl had not taken the drug, would she have developed a clot? This point is argued in the article by Chris Longhurst: “At the end of the day, we don’t know whether it was the right decision.” This ignorance puts us in a dangerous territory, making it a challenge to tell apart the prescient from the charlatan.

• The Big Data world is filled with "events data". You have a log of everyone who clicked on a particular button, or a log of everyone who called your call center, etc. You only have the cases but not any non-cases (e.g. the unhappy customer who did not call the call center). Heartwarming stories like the girl's avoidance of clotting get repeated (or become viral, in modern terminology) but stories of failure are not usually deemed worth reporting. The following table shows four possible stories:

The media imposes a filter so that only the one story will get through. Without mentally accounting for the other stories, one can't judge how important the reported story is!

***

In the July issue of Significance, the magazine by RSS and ASA, Julian Champkin contributed a great profile of Iain Chalmers, the founder of the Cochrane Collaboration, the organization that aggregates and summarizes trial results. I saw this fantastic quote, which speaks to the New York Times article:

Dr. Spock’s 1946 book, Baby and Child Care, was ... read by a huge proportion of [parents around the world]; throughout its first 52 years in print, it outsold every other book except the Bible. “It recommended that babies should be laid to sleep on their stomachs. Now we know that doing that increases the risk of cot (crib) death. Tens of thousands of babies died needlessly because of that advice.”

 

Experts clinging on to their predictions

Andrew Gelman touches on one of my favorite topics: prediction accuracy, and experts who cling to their predictions. Here's Andrew at the Monkey Cage blog.

His starting point is a piece by sociologist Jay Livingston on how various well-known economists made vague predictions (e.g. "I see inflation around the corner") and kept clinging to them (eventually, there will be inflation).

Several theories are given to explain this behavior. One is the idea going back to Kuhn on how scientists stick to their beliefs in the face of negative evidence. Another is that the top economists have invested a lot in those now-questioned beliefs. A third idea is that when the facts and the theories collide, it is cognitively easier to manipulate the facts than to try to change one's theories.

Andrew speculates that "the cost of being wrong is less than the cost of admitting you were wrong." I think there is a certain truth to it. What if you are a famous economist whom people in the field look up to, and frequently cite in their own arguments? You know that no matter what you say, your followers will repeat, and if it's jarringly wrong, they will just keep their mouths shut, then you are under no pressure to correct yourself.

Back to the question of predictive accuracy. I read recently a press release by Microsoft in which they boasted that they predicted the NO vote in the Scottish independence referendum. They said they issued something like an 80 percent chance of NO a few days (or a week) prior to the vote. And since the result was a NO, they were proven correct! What if I had made a 60-percent NO prediction? I probably would have declared victory too.

When we make a statement about predictive accuracy, there has got to be a formula for determining what it means to be accurate. For one-time events like the NO vote, it seems hard to come up with a formula. But without such a formula, you shouldn't believe anyone's claim of accuracy!

Princeton's loss of nerve

I have earlier reported that Princeton's new President has initiated a review of their "grade deflation" policy that was put in almost ten years ago. As you may recall (link), grading in U.S. colleges has become a farce: at top-tier schools, getting an A means you are an average student; not getting an A is many times more informative than getting an A.

The new administration at Princeton has now decided to abandon this cause, pending a faculty vote next month, and return to the old normal. This development is highly regrettable, and a failure of leadership. (The new policy leaves it to individual departments to do whatever they want.)

The recent Alumni publication has two articles about this topic, one penned by President Eisgruber himself. I'm not impressed by the level of reasoning and logic displayed here.

***

Eisgruber's piece is accompanied with a photo, captioned thus:

The goal of Princeton's grading policy is to provide students with meaningful feedback on their performance in courses and independent work.

Such a goal is far too vague to be practical. But let's take this vague policy at face value. How "meaningful" is this feedback when 40% of grades handed out are As, and 80% of grades are either As or Bs? (At Stanford, Harvard, etc., the distributions are even more skewed.)

This tortured logic seems to suggest that the deflated grades are somehow less "meaningful" than the inflated grades of the past. How so? (The grade-deflation policy establishes guidance to limit As to about 35%, apparently by department but also not strictly enforced.)

Eisgruber offered the following justification for rethinking the grade deflation policy:

Almost 10 years after its enactment, the policy remained a lightning rod of controversy and a considerable source of stress for many students, parents, alumni, and faculty members. And regrettably, none of our immediate peer institutions followed our example in taking tough measures to address grade inflation. As a result, Princeton, which ought to be renowned for the unsurpassed quality of its teaching, was attracting more attention for the severity of its curve.

Again:

The committee found no evidence that the grading policy hindered Princeton students' competitiveness in seeking postgraduate employment, fellowships, ... Perceptions of the policy, however, have been a very real source of stress for students, which concerned the committee.

None of this has anything to do with the meaning of grades. In fact, if stress is the primary concern, one might do away with grades altogether.

***

The committeee also did its best to contort the interpretation of data. This is the chart showing the extent of grade inflation at Princeton (kudos to the administrators for making it public - looking to you Stanford, Harvard, Yale for the conspicuous silence):

A reasonable conclusion from the above chart is that the creation and enactment of the grade deflation policy has led to a shift of the distribution of grades from As to Bs. Another clear message is that the 35% target was not robustly applied as even after the policy came into place, the proportion of As stayed above 40 percent. (I am not sure how to explain the apparent plunge in students taking Pass/D/Fail courses around the same time. I'm guessing a separate policy change occurred at that time.)

Worryingly, this is not how Princeton views the chart. The committee members focused on the fact that the shift started "in advance of current policy's implementation in 2005", therefore "the numerical targets may have been only partially responsible for reversing a pattern of higher grading". They then claimed that this other cause is "sustained conversation about grading." Without any data, as far as I can tell, not only did they claim the existence of such an effect but they also asserted that this factor "may be as effective as numerical targets in keeping rising grades in check". In other words, based on the chart, they concluded that the change in proportion of As is caused by two equally important factors, the explicit policy to curb As, and "sustained conversation about grading".

In fact, the committee members seemed confused as they also said "the fraction of A grades... increased between 2009 and 2013 as monitoring of the policy grew more lax." What happened to the sustained conversation?

***

There are two central contradictions between the diagnosis and the proposed solution.

The committee made a case that changing grading policy requires cooperation. If only Princeton takes the lead, and no other peer institution follows, then Princeton is at a disadvantage. And yet, they are proposing to dismantle a coordinated school-wide policy in favor of each department doing its own thing. I'd argue that at the department level, the same dynamic applies. There is no incentive for any particular department to take the lead in curbing As. In fact, the natural equilibrium is for all departments to grade inflate. (Think prisoner's dilemma.)

The other concern of the committee is the "stress" caused to everyone by the new policy. Presumably, such stress is what sustains a conversation about grading. I'd like to know how it is that such conversation would persist in the future.

The final word appears to be a rejection of quantitative measurement. Here's Eisgruber:

The committee wisely said: If it's feedback that we care about, and differentiating between good and better and worse work, that's what we should focus on, not on numbers.

The wisdom has eluded me.

 

 

PS. [11/25/2014] Andrew Gelman and his readers have some good discussion about this post here.