Back in the day, when Amazon was still trying to be the world’s biggest book store, the online retailer worried about overwhelming shoppers with too many choices as they searched for new or additional purchases. To help shoppers, they tried showing them, “Customers who bought this also bought…” But those early attempts at recommendations soon succumbed to what was known as the “Harry Potter problem.” So many customers across the demographic spectrum had already bought the first Harry Potter book, it invariably ended up as the top recommendation. A then-young Amazon developer named Greg Linden decided to take matters into his own hands. He created an algorithm to factor in browsing behavior as well as purchase history, and the online retailer could finally give shoppers Wizard-free recommendations. Linden’s magic did not go unnoticed by upper management. None other than Amazon CEO Jeff Bezos crawled into Linden’s cubicle, bowing on his hands and knees to thank him, uttering, “I’m not worthy. I’m not worthy.”
As it turns out, search is no less important to big STEMM meetings than it is to e-commerce—especially if those conferences include massive numbers of posters. When the Society for Neuroscience convened in Washington, D.C., for its 2017 annual meeting, attendees had to choose from among 13,000 poster presentations plus another 500 sessions over the course of five days. While this wealth of content speaks to the society’s prodigious growth in the last decade, in years past, it has also made some compulsive neuroscientists acutely anxious as they searched for sessions to attend—what social scientists call FOMO (Fear of Missing Out). One of those experiencing FOMO was Dr. Konrad Kording. “The conference had gotten so unbelievably big,” he says, “I couldn’t be sure I was finding all the things I was interested in by using the standard search, and that was making me very unhappy. I figured that there must be a better way.”
Rather than just complain, Kording decided to create a better way. Working with his Bayesian Behavior lab at Northwestern University (he currently runs Kording Lab at the University of Pennsylvania), Kording and his team created a new type of scholarly search which they called, “Science Concierge.” They tested it in 2015 to see how it compared with traditional results and, in 2016, CTI Meeting Technology, the abstract management vendor for SfN, incorporated this new spin on a recommendation engine into the Neuroscience Meeting Planner. In 2017, more than 20,000 items went from the planner’s Recommendations tab to the attendees’ itineraries. Anecdotally, nothing may have spoken more to the Planner’s matchmaking success than the Washington Convention Center’s chronically congested exhibit hall for the poster sessions. Even though it sprawled longer than a football field, it was rare to find that forlorn presenter, standing alone in front of a poster easel.
Kyle Hayden, SfN’s director of meeting programs and attendee services, says her feedback on the recommendations is “very positive”—especially for the way they work in concert with the planner and mobile app (which is now used by more than half of those attending the conference). “While we are providing a large number of things to do, it can make the meetings hard to navigate, so the recommendation engine is a good tool to narrow down what interests attendees most. We also see fewer requests for printed programs, which are hundreds of pages long.” Of course, SfN would rather not print more programs if they don’t have to and no doubt attendees would rather not have to lug them.
But for Mark Coe, CTI’s CEO, the planner’s recommendation engine is about more than convenience. He says, “Ironically, as these meetings grow bigger, there is increased pressure on our clients to make them more personalized, and somehow this must be done efficiently—what’s called, ‘Mass Personalization.’ I know that sounds like an oxymoron if you’re not a techie, but Konrad’s algorithms are a good example of how we can accomplish that goal.”•
These days, thanks to Amazon and other online stalwarts, the recommendation engine is by no means a foreign concept. But there’s a significant difference between the e-commerce approach and the Science Concierge. The retail concepts suggest additional purchases based on past purchase and browse history. Kording and his associates did not find this retrospective perspective suited to the late-breaking, fast-paced nature of scientific research—especially the research presented at conferences. Instead, they opted for a system that would build on the abstracts first added to the attendee’s itinerary. “I know what I like,” Kording says, “so that’s what should be most important in recommending other papers.”
But Kording’s lab did not use the typical search tools for finding those other papers—namely keywords and topic categories. “Keywords are the worst,” he explains. “They are chosen by the authors and not necessarily in a uniform way. They are also usually selected from a list provided by the conference that is nowhere near complete.”
Rather than use a predetermined list of categories, Science Concierge bases its comparisons on virtually the entire database of words contained in the abstracts selected for the conference. Using Latent Semantic Analysis, it breaks down every abstract into phrase clusters and then compares it with other abstracts by analyzing how many clusters they have in common. When a recommended abstract is added to an itinerary, it increases the number of clusters to compare and pulls up even more relevant presentations. As a “sanity check,” Kording’s lab found that the Science Concierge recommendations fit inside SfN’s human-curated categories, but had similarities even more granular than those represented by the conference’s 500 topics. At times, the engine’s results can also pull together abstracts that straddle categories—something that CTI’s Coe finds especially helpful for neuroscience, which has become increasingly multi-disciplinary. “If you look at SfN’s program for 2017,” he says, “you see fifteen to eighteen subtopics that could each be their own meeting.”
Beyond what Science Concierge can do to help SfN, Kording also sees it battling what social scientists call, the “Matthew effect.” The term refers to the New Testament gospel according to Matthew, which laments how the rich always get richer. Kording explains, “When you go to a very big meeting, you have such an intense fear that you may be missing out, you choose a session because you see a famous person is making a presentation–not because of the research that person is presenting.” In other words, the well-known scientist attracts that many more attendees whose time would be better spent elsewhere. Kording believes his algorithms are an antidote to that syndrome. “With the recommendation engine,” he says, “you lose that fear of missing out because you can find people working on the science you most care about. Then you can learn, talk later and maybe collaborate on something in the future. That makes for better science, and it’s what meetings should be about.”
To read the full article, visit CTI's website here.