The Twitter Ranking Algorithm | Hacker News Reader

The Twitter Ranking Algorithm

dl.acm.org

31 pointsJonathanBeuys13 comments

Still inferior to my algorithm: populate an array of tweets by people you follow and sort chronologically with various customizable filters. My paper is still pending, but it will show that the technique leads to a large decrease in fake news, insurrections, polarization, and rage on the internet

1000%, recommendation algorithms inevitably maximise for emotional reactions, promoting polarisation and negativity.

The challenge with publishing the algorithm in the hopes of increased transparency and a broader understanding is that it is still impenetrably complicated to the average user.

Even with an open source algorithm the results on the feed for an individual may still seem opaque based on the input data and the metadata for that input data (not all of which would necessarily be visible).

Wouldn't it be better to have some explanation along the lines of "Why am I seeing this ad?" which is presumably more immediately useful to a larger audience.

Yep, it's gonna just be like when they "open-sourced" their image cropping algorithm and it's just that they provided the model and didn't specify their training data. Granted, yes, you can provide analysis like they did on Github [1] but you don't know _why_ it's outputting something.

[1]: https://github.com/twitter-research/image-crop-analysis

Incredible - Musk wins again by getting the Twitter Algorithm published 18 months ago.

Indeed. He managed to disrupt another industry, this time the fabric of space and time! What a time to be alive.

you can't trust the description of an algorithm, unless you can examine its implementation.

The promise was to "open source the twitter algo" not have a room with glass walls so you can watch the algorithm run line by line, slowly enough to be checked, for every single person that refreshes their feed.

that's exactly what i wrote: it's just a description of the algorithm in a paper, which may be out of date or inexact, unless you can examine the source code

Seeing the source code wouldnt help, you dont know that they are running that code anymore than you know the description was implemented correctly. You would need to see the electrons flying around to know that its running that code.

> Seeing the source code wouldnt help, you dont know that they are running that code anymore than you know the description was implemented correctly

That would be correct if you talk about something entirely data driven, or based on an ML model. For example you could have an entirely table driven algorithm, where the table is read from some configuration file. You could then refrain from publishing the config file. However I don't think that Musk would pull such a trick. Or more exactly: I hope that this is not his intent.

I hope that he wants to build the twitter brand by establishing trust on the part of twitters end users; a move like this would not establish any trust and would make his acquisition worthless. (that may be a naive interpretation, don't know)

> However I don't think that Musk would pull such a trick.

Reading the tea leaves, the CEO under Musk will be the same person who was CEO till 6 months ago, Jack Dorsey.

But it will be totally different this time.

Really? Could you please share your tea leaves? That would be a fascinating turn of events...

Reply on news.ycombinator.com