Ben London

the most important thing about 1993 is that it was an amazing year for music

What this program needs more of, IMO, is people committed to empirically grounded theory.

Offline RecSys evaluation is notoriously hard to get right. Unobserved confounders are often overlooked, but make it even harder. Check out the teaser for my @CONSEQUENCES_ws talk at #RecSys2023 w/ @BenjoLondon! @ACMRecSys @sharechatapp @AmazonScience youtu.be/kbxhZE6Gwto?si…

@roydanroy Nina Mishra (not sure if she’s on Twitter)

Who's an expert in anomaly detection out there? What's are the key texts in this area? Does it have a rigorous theoretical foundation? That handles the adversarial aspects? And what about deep learning and in-roads there?

@haldaume3 Similar thoughts re: LLMs for recommendation. For instance, they can generate reasonable looking playlists based on a generic theme (eg “alt rock”)—which is incredible!—but may struggle with niche queries, and the playlists are usually kind of boring.

Which of the following words do you think is most rare in US English? Can you give me a *rarer word* than these (try it, then look it up on google ngram viewer). And what does this have to do with large language models like ChatGPT, BARD, etc.? Read on... >

Damn iPhone keyboard Makes all my text messages Come out as haikus

GO BIRDS! 🦅

@HudsonKate You’re pudding me on

Thank goodness iOS can edit sent messages! Now I can stop texting about all these ducks and shot.

Open AI chatbot prompt: "Rewrite Baby Got Back in the Style of the Canterbury Tales." Once upon a time in the land of Canterbury A lady fair with a behind so hearty Walked in the room with a waist so small And a round thing in front of us all ...

@tommy_da_cat They’re all over on the “Furry” server

Ok you lost me already #Mastodon #academia

@liambaranauskas Bet his mouth is full of joyful turds tonight!

The guy who ate dogshit after the Eagles won the super bowl should be higher in the pantheon of frequently-referenced Philadelphia sports fan behavior. I can't think of anything more Philadelphian than being so happy you decide to eat dogshit

@DimitrisPapail Would an example f be SGD? Like, you’re trying to learn the SGD update?

Say I want to infer a function f() from its evals. That's hard! But what if f(x) is an "algorithm" comprising T iterations, i.e., f(x) = f_T(...(f_2(f_1(x))) and f_i are "simple" (e.g., low-deg poly?). Do we know how do to recover these f_i's?

@roydanroy @DimitrisPapail Or Rademacher complexity. With that, if each f_t is Lipschitz, you can recursively apply the contraction lemma, for which there’s a vector-valued version. But all the Lipschitz coeffs would multiply, so you’d hope they’re <1 so you don’t get O(C^T).

@DimitrisPapail There are algebraic ways of bounding VC dimension (good for binary valued functions) for compositional classes. I suspect the same is true for functions take multiple values.

Today at 12:05 @CONSEQUENCES_ws

@olivierjeunen And now I know the workshop handle :)

As I only just found your Twitter handle: @BenjoLondon :)

Looking forward to this one! #RecSys2022 @CONSEQUENCES_ws

Collaboration with Giuseppe Benedetto & Thorsten Joachim, who might not be on Twitter

Fri, 9/23 @ 12:05 PDT: @AlexanderBuchh3 will present our work on a new estimator for offline evaluation of ranking policies. amazon.science/publications/o…

Presenting 2 papers on new work at CONSEQUENCES+REVEAL Workshop @ RecSys ‘22 sites.google.com/view/consequen… Details 👇