for the sake of society and the future of humanity, recommender systems should introduce an element of randomness ... say I view items X and Y, and I am only very slightly inclined to pick Y .. next time the system shows me Y1 and Y2 because I like Y's... I pick Y2 and then it recommends Y21 and Y22 etc. .. never again would I see X, X1, X2, etc. although I might have been interested in checking them out, it's just that I had to pick something

I really think Amazon, Youtube, Netflix etc. shpuld have a "random picks" section .. not bestsellers, not most watched, just random

The short response to your comment is: there sometimes sort of is, but it's hard.

The way your implying how recommendation systems work seems like a tree structure, where if you start down one path of items it precludes you from another path. This isn't quite how things usually work (at least in my experience).

Typically recommendations are calculated using sparse matrices with users along one axis and items along another. This would make cell values a user's affinity for the item. You can then use ALS algorithms to help compute missing values and predict a user's affitinity. There's a lot I'm glossing over here, so if you're interested I recommend reading further on this.

One of the trickiest parts of any recommendation system is a cold start, what to do with a new user. Sometimes randomness does play a part here, but more often it's based on popularity or recency. And sophisticated recommendation systems will combine those along with data from other 3rd party sources.

(continued...)
Additionally, you will always get that asshole exec that asks "why am I being recommended this piece of content?". To which the answer should never be "I don't know". So randomness doesn't help much there. In my experience, randomness has played most favorably in the ordering of items that come back, but that still makes it tougher to explain.

(source: I worked on recommendation systems for several large cable/satellite TV companies for several years)

@afreitas I didn't mean to imply a tree structure, maybe the naming I used was confusing .. in my example Y1 & Y2 are meant to point to something recommended based on my choice of Y, regardless what the algoruthm used was, as long as it is deterministic. i.e. Y could be a cooking book and Y1 a travel guide, but it was recommended because users who bought Y usually bought Y1 and because Y1 is the most popular travel book in ur region or whatever

the point is, the system assumes that people know what they want/like

@afreitas actually, thinking about what you said again I guess the bit I am assuming is that a recommendation is always affected by (or should I say based on) a user's past choices .. is this a correct assumption?

regarding the bit about someone asking y they were recommended a random item, what I am thinking is a clearly labelled "random items" sections in addition to the recommendations

I always feel recommendation systems are weighted but they do tend to sink down to what you most recently viewed so very much like a choose your adventure . The goal is to get you to perform an action and sooner they can narrow your options the sooner you buy shit and avoid being paralysed by choice. Along the way i guess next visit can take into account you viewed x for a long period but y little yet you bought y let us show you more x.

As for exploring styles like youtube. Utter randomness would seem like spam if it has no relation to previews things.