(this is a bit of a rant, i’m sorry)

what in particular do you mean by lack of discoverability?

like, i want to see posts from communities that i already subscribed to, but because there’s more than 1000 communities on the fediverse and i’m only subscribed to a small countable subset of them, i inevitably lose out on a lot of content. (The “all” feed sucks unfortunately). So how to solve this?

The lack of discoverability is non-starter for many.

The Fediverse significantly lacks behind on the Content Discoverability technology.

I guess this is because there was a loud public outcry in the last 20 years that whoever makes your feed (this is called an “recommendation algorithm” or abbreviated “the algorithm”) has a lot of political power to decide what you see and what you don’t see, and that’s frowned upon. Because everybody that has power over what you see and what you don’t see is bad. That is why nobody wanted to provide an recommendation algorithm for the fediverse, because they would expose themselves to wild accusations. There should be an open-source recommendation algorithm, though; I’m sure of it.

  • Ephera@lemmy.mlEnglish
    2·
    21 hours ago

    There should be an open-source recommendation algorithm, though; I’m sure of it.

    Problem is that the kind of algorithm you envision is technologically a black-box, not just by choice. It’s a machine learning model. At best, you could make the training data and instructions public, but it would still be hard to reason why it makes certain decisions. Corporations traditionally try to eliminate biases by throwing as much data at it as possible, but that makes it even harder to reason about it.

    I guess, maybe you could try to split the tasks. So, set up a list of e.g. 50 topics, such as sports, IT, politics etc… Then use a small language model to decide into which categories each post fits. And then you could let the user decide the weights for the topics + weights for recency and vote count.
    Or I guess, automatically decide the weights based on what the user upvotes and then make the weights transparent to each user.

    But yeah, I don’t think there’s prior art in this respect, so would probably need lots of experimenting still.

    • gandalf_der_12te@discuss.tchncs.deOPEnglish
      1·
      20 hours ago

      hmm, i think you’re overthinking this. what if the recommendation algorithm simply gives you stuff from communities and you’ve subscribed to and “similar” communities (these would have to be linked from the original communities / link to the original communities)?

      that should be reasonably easy and not involve any neural networks. i think basically it constructs a “feed” (post list) which is basically a remix of other lists (which are the individual communities that stuff is taken from), maybe weighted with a certain scalar factor.