Active Stocks
Thu Mar 28 2024 15:59:33
  1. Tata Steel share price
  2. 155.90 2.00%
  1. ICICI Bank share price
  2. 1,095.75 1.08%
  1. HDFC Bank share price
  2. 1,448.20 0.52%
  1. ITC share price
  2. 428.55 0.13%
  1. Power Grid Corporation Of India share price
  2. 277.05 2.21%
Business News/ Opinion / Columns/  In which I learn from bras
BackBack

In which I learn from bras

We live in an age when algorithms can quickly sift through data to find patterns

Photo: iStockPremium
Photo: iStock

What’s to be done, when every webpage you visit is overrun by ads featuring lissome feminine torsos wearing bras? That’s my fate of late and I promise you two things. One, I did nothing to prompt these targeted ads; I think someone in the family must have recently used my laptop to shop for bras. Two, I’m not complaining.

Still, think about what happened here. There are algorithms in place somewhere in Google-land that monitor your web activity. Based on some patterns they detect, they offer you specific suggestions for future browsing, or spring these specific ads at you. It might be something as simple as the “customers who viewed this item also viewed" list on Amazon. Or it might be the deduction that because you pored over a page about the Caribbean, you’re interested in cruises, and here are some for you to consider (this actually happened to me). Or of course, it might be bras.

The point is, we live in an age when algorithms like these can sift quickly through reams of data and find patterns. Of those, some patterns might mean something, others nothing. It might annoy you that this happens at all, meaning that your movements across the web are monitored and mined for possible gain. But such algorithms are a nearly inescapable reality now. In effect, they are learning from the behaviour of millions of internet users, from the vast amounts of data generated as they simply click their way around the web.

And that should suggest other places we can use learning algorithms like these, their results possibly more relevant and less distracting than bras. What I’m getting at here is covid, and particularly, testing for the virus. Let’s say you own a movie theatre that’s just reopened its doors after many pandemic months. You’re desperate for customers, but you’re also aware that the pandemic is not yet a thing of the past. So you decide that you will admit only those patrons who pass a test for the virus, administered at the entrance.

Question: Do you test everyone who turns up to watch a film? Well, however quick the test is, this strategy is likely to produce a long line of disgruntled patrons waiting to be tested, a bottleneck that will turn many of them away from your theatre. Can’t have that. Besides, given the prevalence of the virus, the majority of those tests will likely return negative results. These covid tests are not cheap, so it makes very little economic sense to willingly take on this expensive futility.

But if you don’t test everyone, whom do you test?

Maybe you don’t think we need to pay much attention to movie theatres as we emerge from the pandemic. Maybe there are other areas of the economy we should care about. Fine, so consider a country that’s trying, after many pandemic months, to open up to tourists. What if tourism is a major part of its economy? Clearly, it is now desperate to bring in tourists again. Yet just as clearly, it cannot afford to let the virus spread again. So the country has to be careful about who is allowed to enter. 32 million tourists visited in 2019, the last year before corona. Assuming numbers will return to that kind of level, can this country really test every tourist who wants to enter?

That is: Can Greece afford the time spent, the money involved, the annoyance caused?

Short answer: no. So is there some way learning algorithms can help? Can we use them to identify, based on data that is available about people seeking to enter Greece, who is likely to need testing?

Short answer: yes. In fact, last year, Greek authorities put in place a “reinforcement learning system nicknamed Eva" on all its border entry points. Its aim was to “limit the influx of asymptomatic travellers infected with" the virus. It used demographic information—sex, age, country of origin and more—for incoming tourists, extracted from a form that they are required to complete a day before arrival. It also used “testing results from previous travellers." Those quotes are from a recently-published paper by some scientists who studied the problem and then put in place an algorithm to choose tourists to test (“Efficient and targeted COVID-19 border testing via reinforcement learning", Hamsa Bastani, Kimon Drakopoulos, Vishal Gupta and others, Nature, 22 September 2021; go.nature.com/3lZgklJ).

There are some intriguing details here. From the demographic information, Eva extracted a set of “discrete, interpretable traveller types", though those types were updated every week with results from recent and ongoing testing. The algorithm then used an “empirical Bayes method" to estimate the prevalence of each of these types. This was important because the data is otherwise flawed in two ways: there are only a few positive cases among people tested (“imbalanced" data), and there are only a few arrivals from some countries (“sparse" data). The Bayes method counters these limitations.

Going beyond that, Eva seeks to achieve two things with its choice of who should be tested. One, identify as many infected but asymptomatic tourists as possible. That’s an obvious goal, of course. But two, which is not so obvious, also test traveller types “for which it does not currently have precise estimates"—so as to learn from those as well. The scientists emphasize that this latter step is a vital one in Eva’s education. For let’s suppose Eva only tested people from apparently high-risk groups. In just days, then, it would not have any “recent testing data about many other [moderate-risk groups]." This is a serious flaw, because all over the world we’ve seen covid cases rise sharply, suddenly and swiftly—again and again.

Take India’s second wave in April this year, for just one example. Daily case counts shot up from about 50,000 at the beginning of that month to over 200,000 by mid-April, to nearly 400,000 by the end of the month. If there were Indians visiting Greece at the start of April, and if Eva wasn’t testing them because citizens of other countries were seen as greater risks, then it would likely miss that calamitous rise altogether.

With all this and plenty more in its design, Eva delivered impressively. The scientists estimated that during the peak summer tourist season in Greece, random testing would have identified just over half of the infected travellers that Eva’s targeted testing did. Put another way, if random testing were to be as effective as Eva at identifying those infected, Greek authorities would have had to administer “85% more tests at each point of entry" into the country. You can imagine, therefore, the cost savings due to Eva.

Interestingly, Eva’s performance in relation to random testing dropped in the tourist off-season. This should not be a surprise, for with less tourists, the fraction tested increased, thus “reducing the value of targeted testing." To understand this, think of testing everyone who enters Greece. Precisely because everyone gets tested, Eva offers no benefit.

The point, then, is that Eva is most effective when tests are scarce relative to the incoming tourist numbers. Which is exactly how it should be, of course. It’s when tourists are flooding in that you want the most efficient mechanism possible to test them and identify who’s infected.

It’s when people are looking for bras that you want to throw ads at them and urge them to buy. Again, not that I’m complaining.

Once a computer scientist, Dilip D’Souza now lives in Mumbai and writes for his dinners. His Twitter handle is @DeathEndsFun

Unlock a world of Benefits! From insightful newsletters to real-time stock tracking, breaking news and a personalized newsfeed – it's all here, just a click away! Login Now!

Catch all the Business News, Market News, Breaking News Events and Latest News Updates on Live Mint. Download The Mint News App to get Daily Market Updates.
More Less
Published: 21 Oct 2021, 09:48 PM IST
Next Story footLogo
Recommended For You
Switch to the Mint app for fast and personalized news - Get App