It is not solved. Let me go into a bit of details: the current state of the art is using two-tower like models that create dense embeddings both for users and items. These towers are, in essence, two models that need to keep into account histories, reviews, pictures, color, material, ... . So this problem is as solved as the models that can handle these problems are.

Currently, I would expect that teams are experimenting with repurposing Attention-like architectures in some way to get better embeddings, especially from sequence like features.

Do you feel that recommendations given to you are perfect? That more or less should answer the question.

Evaluating recommendation systems is hard because you actually require a human in the loop. Even worse, giving the recommendations alters the human behavior. Then you need to think what metric are you going to use. For training you will most probably use a proxy metric that correlates. Maybe you want to optimize different metrics and they actually need to be balanced. Then there are lot of confounding variables: maybe a better UX will improve the metrics than a better algorithm, or a change of products.

It seems that with big enough data you can improve old models with deep learning but I think recommenders are very far from similar gains to other fields (NLP or CV for example). And most companies don't have that much data.

Rec systems that tech giants use looks vastly different than what ML academics study. Industry rec systems are a constantly evolving beast. Why? Because these teams of ML engineers are constantly running A/B tests to tweak some aspect of the model. The resulting model likely doesnt fit nicely into something category of model you can describe as collaborative filtering/lambdamart/etc b/c it's some highly performant glued together mess.

You'll find a detailed answer on paperswithcode, selecting the "Recommendation task":

https://paperswithcode.com/datasets?task=recommendation-syst...

Like many have already said, this is mostly an academic answer (even if some of the papers are written by the industry).

In the industrial world, the answer is a lot more subtle. Each domain has its own constraints and the best method will vary.

Also, keep in mind that in some domains like ad tech, the whole measurement process is messed up by the attribution mechanism, which attributes a sale to the last click (which clearly is a poor indication of whether the recommendation engine is doing a good job -- the industry settled on this attribution mechanism because it is easy to audit).

There's more to explore. In particular, using groups of similar products and users to infer recommendations given only very sparse data is powerful but underexplored in production usage.

(Disclosure: I work at a a company which just expanded their product portfolio in this direction. The first few pilot customers show very good results, but there are a lot of aspects and nuances and alternative approaches we haven't had time to try yet.)

Recommendation will never be solved (like nuclear fusion can be) because it is about human minds liking or disliking things, and so it is pretty much "psychology-complete".

You will only see the field make steady, regular and infinite progress towards an unknown asymptote.

And how does it perform compared to the solution that won the Netflix prize, years ago?

>Just curious, what is the state of the art of recommendation systems these days?

Asking your friends for recommendations. In my opinion automated systems are still not nearly as good as another person knowing you and your preferences.

One area to explore is how to optimize a sequence of events (e.g., user journey). Reinforcement learning (RL) might be useful for that.

There's always some new neural network module that pushes the needle.