mastodon 4.7.3I build recommender systems that work for actual people. Aspirational cyclist. Actual dog person. Suspiciously majestic. Not even remotely neurotypical.
I run RecSys.social. I don’t accept follow requests from accounts without a bio or profile picture.
This is your annual reminder that many autistic people consider groups seeking to prevent or cure autism to be eugenicist hate groups and would strongly prefer that any donations you make go to groups that seek to improve the lives of autistic people instead
tfw you’re reading an article in Communications Of The ACM and surprised to see own your name
https://cacm.acm.org/research/a-task-centric-perspective-on-recommender-systems/
the mass production of knowledge-shaped objects is insufficient to resolve design problems without correct answers
breaded pesto pork chops with roasted potatoes, green beans, and jammy tomatoes 😋
Twenty years later, I finally parted with my college notes from a degree that launched my career in a sense but was in a completely different field. Letting go of who you used to be is hard sometimes.
Adventures in #RecSys:
Model A can’t recommend Alien, but Model B can’t recommend Batman. Which one is better/worse? What on earth is happening here? Why can’t we find any papers about this?
“The most important discomfort that powerful people experience is having ego-shattering conflicts with subordinates who know how to do things they do not know how to do.”
short version of recent literature on #RecSys evaluation: this must be a sketch comedy show, because apparently everything is made up and the points don’t matter
after a lot of time spent trying to mitigate the downsides of the BPR loss function (i.e. popularity bias), my recommendation continues to be “use something else” but…if you must, consider mixed negative sampling from both the uniform and item popularity distributions
C: “the main competency of the kaiju is to crash into large man-made structures and knock them over! a wall is not going to stop them”
“I need chopped cilantro, stat”
I dunno if anyone will pay for little essays and experiments about recommender systems, but I sure am having fun coming up with them
another day reading the RecSys literature and walking away with the distinct impression that we don’t understand much of what we’re doing (even when it works)
I’ve said for a long time now that if you average item embeddings to make recommendations, items similar to the average embedding might be either a semantic blend or something wildly unrelated (because learned embeddings don’t provide global semantic smoothness, only within local neighborhoods.)
Today I found a new failure mode: the average of Great British Bake-off with anything is still just GBBO.
My new RecSys toolkit is too powerful for public release; the recommendations broke containment during testing and started recommending themselves to people who hadn’t asked for but desperately needed them. I’ve decided not to make the toolkit generally available and instead will be using it as part of a defensive RecSys program with a limited set of partners.
this project has now reached the “whiteboard in the living room with fancy colorful markers” stage, so I fear there’s no turning back now
It’s pretty incredible that I can get better “similar TV show” recommendations out of a text embedding model and some low-effort DuckDB queries than I can from many actual streaming services
Reading newsletters about recommender systems, I’m having conflicting experiences:
- There are some really smart people writing some great stuff that distills papers and clarifies trends in the field! 🤓
- It’s not obvious how I’d use any of this to actually produce better recommendations that real people appreciate and enjoy 🧐
So I tried out the new-ish EVoC clustering algorithm, which apparently is specialized for dealing with embeddings, doesn’t require much tuning, and should produce good results out of the box.
It didn’t adding cluster ids to most of the embeddings. Tried suggestions from troubleshooting guide, still didn’t put most of them in clusters.
Returning that one back to the shelf.
someday that statue is going to be the subject of a very famous photo
Cheese a killer cleaned
Clam chowder, mondegreen
Fly-by-night with a buttercream
Dragon speed to the overmind
(Anaheim)
Comprehended saffron rice
Sensational, in Fahrenheit
Personified?
I suspect this afflicts many of us who work in technology these days:
What should I watch (at home) if I liked these movies?
- The Martian
- Moneyball
- The Big Short
- Enemy At The Gates
After poking around with basic text embeddings of TV show metadata (quite a lot), the only failure cases I’ve identified in similar item recommendations have more to do with missing or uninformative metadata than they do with any kind of technical issue.
The biggest downside is that these embeddings don’t encode anything about popularity or perceived quality, which results in some obscure recs. Easy to fix with a minor amount of score boosting though.
not sure if “slowly feeding cute girls string cheese” was on my bucket list, but she insists I should check it off regularly
“we use a 2D sphere”
sir, I believe that is called a circle
I don’t care if the performance estimator produces theoretically unbiased metrics unless you’ve actually looked at your own recommendations and honestly assessed how genuinely useful and/or laughably bad you find them
“she might as well be picking a fight with the composer”
oh good, you’ve understood correctly
Saw a paper that called it “Bayesian Pairwise Ranking” and given the degree of popularity bias, that is definitely more accurate than “Personalized Ranking”
LinkedIn heckling me after declining a connection:
“I don’t know Jack”
