Monteluna
@monteluna
@vrypan.eth hey have you thought about using some tf-idf metrics when generating casts for wdim? Because I'm using "$tipn" in my casts a lot, the algorithm seemed to bias towards casts about "$tipn". Maybe filtering cash tags would be good? Could be a one off error but it basically didn't generate a lot of casts I would find interesting.
1 reply
0 recast
1 reaction
vrypan |--o--|
@vrypan.eth
The current feed is 100% algorithmic. 1. It picks every cast (including replies) from the people you follow. 2. If the cast is a reply, it finds the root cast of the thread. 3. It scores the people you follow based on your last 1000 interactions with them (likes, recasts). 4. It sorts the list of casts generated in 1&2, based on the score in 3. I'm working on an alternative, using AI.
2 replies
0 recast
1 reaction
Monteluna
@monteluna
Wondering if also adding a @quotient score could help. I'm on the android client and can't show a screenshot unfortunately, but my top of feed was basically $tipn folks I've never heard of. Ranking by reactions is somewhat dangerous since it can be bottled, but weighing by pagerank style scorers might improve the algo. You may not have to weigh the reactions themselves but at least the casts themselves might be better.
1 reply
0 recast
1 reaction
Monteluna
@monteluna
@vrypan.eth actually going through those casts, it looks like it followed @ruminations comment casts and that's the signal it picked up on. Not sure if you can test my wdim feed specifically but might be a good edge case to consider. I don't really want to see users I like comment casts to casts of users with a low quotient signal.
0 reply
0 recast
1 reaction
