Personalization is very overrated. I remember reading a big NY Times piece on how their tech team uses an embedding-based personalization and recommender system, with a sexy blog post about their internal team that implements it, and I just recoiled in horror.
For one, searching for content on NY Times site sucks. I just want a better search engine without personalization. Just better categories, tags, search by relevance to a named entity or narrowed in a time frame, which just works. Currently it’s an awful way to look for things. Couldn’t they spend engineering budget better on just search engine problems, rather than some sexy, over-hyped customer embedding model with all kinds of gadgets to solve cold start, etc., but which doesn’t seem to actually add value?
The other thing is that you _know_ the whole purpose is to optimize some notion of engagement that is not actually a healthy measure of the user getting value out of the product.
I think there could be some minimal benefit to stylizing or customizing results based on highly aggregated characteristics, like country or a super broad age group.
But beyond ensuring the user is routed into a very high-level bucket of characteristics that can broadly determine relevance, anything more specific to individuals or anything relying on nearest neighbor or collaborative filtering based on a highly individualized representation, is just flat out unhealthy, by design.
Information retrieval without personalization is like the new version of “I just want my phone to actually function as a phone.”
For one, searching for content on NY Times site sucks. I just want a better search engine without personalization. Just better categories, tags, search by relevance to a named entity or narrowed in a time frame, which just works. Currently it’s an awful way to look for things. Couldn’t they spend engineering budget better on just search engine problems, rather than some sexy, over-hyped customer embedding model with all kinds of gadgets to solve cold start, etc., but which doesn’t seem to actually add value?
The other thing is that you _know_ the whole purpose is to optimize some notion of engagement that is not actually a healthy measure of the user getting value out of the product.
I think there could be some minimal benefit to stylizing or customizing results based on highly aggregated characteristics, like country or a super broad age group.
But beyond ensuring the user is routed into a very high-level bucket of characteristics that can broadly determine relevance, anything more specific to individuals or anything relying on nearest neighbor or collaborative filtering based on a highly individualized representation, is just flat out unhealthy, by design.
Information retrieval without personalization is like the new version of “I just want my phone to actually function as a phone.”