More and more in our life is being Yelpized: our decisions between alternatives guided by scores assigned by unknown individuals. Unsurprisingly, research on this topic is more than sprawling, especially when it comes to the question of how such a system can be designed to be resilient to all sorts of cheating.

After roughly 5 minutes of consecutive thinking, and a weak eco-evo analogy, here is the system I like best:

Replacing scores by trade-offs
(aka a normalized version of the only useful paragraph in your typical internet review: pros and cons)

Artificial boosting is much more difficult if we are not giving a score, but a trade-off between features.

Let's drop the idea that there are better and worse restaurants. Really.

Simply suppose that one might want to recommend - or advise against - a restaurant first for its service, a bit less for its food, not at all for its location. The absolute value is irrelevant; what matters is the ordering -- or more exactly, how it differs from the median ordering (to be robust against the fact that all restaurants are going to be recommended for the food). And now, we can just force the prospective diner to select such an ordering as the key to search for restaurants.

Knowing that a restaurant is mostly recommended for its speed is vastly more useful and robust than knowing it has an average score of 3.7/5. If the owner asks all her cousins to submit reviews, they still have to choose which feature to emphasize - a uniform distribution of recommendations just dilutes the message.

The curious thing is that, given enough time, this system would promote specialization - it would undermine restaurants that so happen to serve delicious food cheap and fast in a dreamy setting. But nothing prevents you from having "well-roundedness" as an emergent feature for search, provided its suspicious nature is made clear to the user, and provided it still matters only in relative terms, i.e. via its position in an ordered collection of features.

Am*z*n headhunters, you can talk to my assistant.(1)

(1) and admire my restraint in not making a Jivaro joke.

Tags: