The journey of a mathematical model that brings together the wisdom of a fisherman, party nights in Berlin and a young AI-based startup.
When I was a kid, I used to go fishing with my grandpa. Getting fish out of water was a mixture of heuristics and chance. We didn’t really know what was happening underwater, nor were we worried that much. It was like a black-box model, for which a certain combination of bait, rope, and thinner or thicker fishing line yielded better results. «There must be a more scientific approach», I said (being a kid, I was maybe not using these exact terms). My grandpa laughed at that and told me not to get distracted, else I’d have lost the catch.
A few years later, I was doing my PhD, spending my days going through mathematical proofs, sometimes sitting for months in front of one group of differential equations. I was researching consensus theory, a fascinating mathematical approach that models how groups come to an agreement. It is used for modelling road traffic, as well as swarms of quadcopters, but it can also explain the way opinions are formed on a social network. A pretty powerful mathematical theory, whose fundamentals are well illustrated in “Information consensus in multivehicle cooperative control”, W. Ren, W. Beard and E. M. Atkins, 2007.
It was during that time that I envisioned a new approach to group recommendations. I was grabbing tea with my friend Amogh, and we were trying to organize a night out with our buddies. An inefficient yet cumbersome procedure: browsing, looking for a restaurant, finding a good bar nearby, thinking if that would fit everybody, proposing, hoping for unanimity, else restarting. «There must be a more scientific approach», I thought, «something that can automate this practice, something that can make us save time».
What if I incorporated consensus theory equations into a deep learning model? An AI automating group decisions, for deciding which movie to watch on Netflix with my partner, which bar to hang out with my friends, or which restaurant to pick with my colleagues, would be a very powerful tool. Yet, I couldn’t find any app that was providing that service: I am an affectionate user of The Spoke and Likewise, super nice and powerful apps, but they are designed for helping friends or similar users share recommendations with each other, rather than giving group recommendations.
«If that doesn’t exist, let’s make it», Amogh told me, with his entrepreneurial spirit. In that very moment, the journey of LifeTap started. We envisioned an artificial intelligence capable of understanding the dynamics of a group, recognizing the preferences of each member, combining those, simulating the dynamics of an agreement, and, finally, yielding a handful of great recommendations. All in a few milliseconds.
«Why hasn’t anybody already created it?» we were wondering. It looked like, when implementing AI, people like to use the fisherman approach of my grandpa: just throw everything inside, add one layer here, remove some nodes there, and hope for the best. This works great when you have many good baits (namely, data) and much time. For giving group recommendations, however, we needed a model, and that model was coming directly from the world I knew well: consensus theory. The open question was how to describe the heterogeneity of a group, or, in simpler words, how to quantify how similar two users are, without the need of employing the old-fashioned collaborative filtering approach.
The answer wasn’t there, ready and available as an off-the-shelf tutorial on Medium.com. We took the chance to have a journey through the nightlife of Berlin, to understand what makes people decide where to go or what to do with friends, as well as what drives groups to opt for one bar rather than for the other. The main goal was understanding if there were a common behavioural pattern beneath, technically, something that could motivate the usage of data for prediction. «On Fridays I like to forget my working week, I need loud music», said a Portuguese girl sitting at the bar. «I agree, but don’t take me to snobbish clubs» echoed the guy next to her.
In another bar, the answers were similar. There was a younger audience there, larger groups, and we started talking to them, asking why they chose that bar. «She decides» one young guy told us, pointing at the friend next to him. «I do not decide… I just suggest, and you simply follow because you’ve no better ideas», she replied. Something started shaping up in my mind: many variables were relevant for the decision-making process, for instance, which weekday, what weather, maybe even the season, the age of people in the group and even their genders; but, mainly, what seemed to have a large impact were the personalities of group members, whether there was a leader (what in the consensus community we usually call “stubborn agent”), and how different personalities interact.
Personality and situational context are the main decision drivers in choosing everyday activities, as presented in “Explaining everyday behaviours and situational context by personality metatraits and higher‐order values”, E. Skimina, J. Cieciuch, 2020. We had found the missing piece of the puzzle, and that allowed us to put the last candle on the cake. Months of experimental validation, collaboration with renowned scientists and entrepreneurs like our advisor Dr. Galen Buckwalter, technical issues, dead-end streets, but, finally, the app incorporating this visionary AI was out and ready.
And in that very moment, remembering those days fishing with my grandpa, I realized that we cannot employ a simple mathematical model for describing everything, sometimes it is just sufficient to use heuristics and hoping for the best. But when using a mathematical model, nice things come out.
