Post

outline tags! pls

Manual tagging by users can be problematic because of deliberate mistagging (I regularly find porn in my 'Flowers' and 'Moss' feeds; was forced to entirely unsubscribe from 'Art'). People will do anything to get seen.

Perhaps I am missing a feature somewhere, but the reason the discover tag is useless to me is because I don't have an option to have it prioritize newness. I get posts from so long ago I'm not at all interested. I would love to be able to weight recency. Next 50 in algo chronologically.

People from digital humanities and comp. social science often use methods based on clustering embeddings (such as BERTopic) as it works well on short texts and scales much better than LDA and LLM-based stuff

Would LLMs do better than ML models here? Too expensive?

That could get messy, I post about several non overlapping subjects. How about adding tags to posts from that analysis as it's written, this would provide a 'this wording gives these vibes" service , could be a valuable tool for understanding you audience and your post. Would need to be optional

+1 for manual user-created signals. Not ideal from a UX standpoint, but yall have a lot of hardcore fans willing to help in order to see the ecosystem succeed. People already do a less useful version of this by eg manually creating lists/feeds. An official Bluesky UI would at least pool user effort.

For what it's worth, I've found standard LDA (implemented in mallet/tomotopy) works quite well on Tweets, despite their length. Then perhaps apply a very small LM to transform the topic words / top documents to human-readable labels?

An ML friend suggested LLMs are better than manual labellers than humans nowadays—and you wouldn't need to use massive LLMs either (but it also wouldn't be free). Also suggested optimising for x*likes + y*reposts + z*replies vs just x*likes (though that might just be misreading what you wrote above)

A model like DistilBERT works well on short text and is computationally efficient - I've trained some models to classify tweets pre-X for instance huggingface.co/docs/transfo...

yeah I think the account and the accounts interacting with it are basically a hard requirement to properly classify, an issue I've noticed w/ *ahem* some art is they either don't tag at all, or *really* like to tag everything except the actual target audience

This really isn’t that hard, you need to look at network structure.

There’s a very algo centric, some what technical detail here but maybe: what is the the design goal?

Please help!!!! I really don't want to see people wearing rubber suits. Today one was in a cage with sexual suggestions!!!! AND I HAVE ADULT CONTENT TURNED OFF. I don't want to have leave Bluesky because of gross content.

not sure if this is already how the social clusters work but graph embeddings of follows/interaction counts would make sense if you want something like a social similarity measurement for ranking

? Unless you are using extremely naive clustering algos almost everything you could do here certainly counts as ML.

Do you need a graph theorist because I. Am. Right. Here. www.linkedin.com/in/maddy-mus...