Comment by zeroq
4 days ago
Great work!
Some five years ago I was day dreaming about recommendation engine for movies where you could say "hey Ciri, give me a good gangster flick", and it will come up with something that you haven't seen yet but you'd definitely love.
To my amazement almost everyone, even true AI believers, thought it was impossible to achieve. :d
But my question is - having such huge dataset, do we really need AI for it? SASRec/RAG is sexy, but could the same result be achieved with simple ranking and intersections like lastfm did in the past with music?
Some twenty years ago I came up with an idea of "brain" data structure for recommendations where you have all your items (books, movies or articles) modeled as a graph, and whenever you pick something it makes a ripple effect, effectively raising scores in cascade of every adjecent item.
Just like your brain works - when you stumble upon something new it immediately brings back memories of similar things from the past. I never had the opportunity to implement it and test in real life scenario, but I'd be surprised if a variant of this is not widely used across different recommendation systems, like Amazon.
> To my amazement almost everyone, even true AI believers, thought it was impossible to achieve.
I mean. Is it possible? Hey Zeroq give me a good gangster flick please?
How will you know what i have seen and what I haven’t yet? How will you know what movies I like? Are there even good gangster flicks i would enjoy and haven’t seen yet?
The way the problem is phrased it sounds like your dream recommender has two properties: “it doesn’t receive any other information than what is in the prompt” and “it always recommends a movie you haven’t seen and will enjoy”. Those together are what makes your dream recommendation engine impossible. If you relax those then of course it is possible. That’s just a recommendation engine.
last.fm used a primitive machine learning algorithm too, else what are you going to rank by?
Did they? I recall similar site back from as far as from 2008. Might be them or something similar.
Anyway. I can totally see such site running purely on statistics. Every song, every artists, every genre is a bucket. You listen to a song you put a drop in these buckets. Once there's enough water running we can compare you to other users and their buckets.
It might be hard to run it on scale in real time, but c'mon, it's leetcode junior level assignment level of complication.
[dead]