Recommendation has been there for quite a long time on e-commerce, ad-targeting, applications for TV, Books, Restaurants, Video Games… And it has been often based on the same pattern: Statistic. Applications using recommendation usually collect and analyse data to extract the best possible outputs toward its users. It’s based on calculation. While it instinctively seems to be the right way to do it, it’s actually not. The reasons are simple: it does not work without initial behaviour data and you can’t take friction into account as it would considerably and infinitely increase the complexity of calculation. you therefore tend to end up with a pile of bias which distorts the results and lead to bad/unrelevant/uncomplete recommendation.