Jacek.degen.eth 🎩 on Warpcast

Content pfp

0 reply

0 recast

0 reaction

Jacek.degen.eth 🎩 pfp

Jacek.degen.eth 🎩

Came up with some metrics for Farcaster profiles. If anybody has experience with ML or similar tech, I would be curious to hear ideas on how to feed casts and profile metadata into an algorithm that identifies spam. https://dune.com/queries/3540715

24 replies

12 recasts

71 reactions

Eddie Wharton pfp

A classic ML approach is releasing labeled data (users marked as spam or not) to a competition. You hold back some data to evaluate submissions Probably good candidate features for spam detection using NLP to parse text in names/bios/casts Maybe a good potential community challenge @ilemi?

2 replies

0 recast

2 reactions

J Hackworth pfp

I talked about this a bit last week but you need to create a supervised training to define what is/isn't spam to train your model. Either you could manually label this is a spam account or what you could do is start with a list of known spam words and then develop additional features on top of that. Happy to talk more

0 reply

0 recast

1 reaction

Ape/rture pfp

I think you mostly want to look at the content of the casts since spam is often defined by the goal of the cast. The cast could shill coins or contain URLs. Also in addition try to identify if the text was human generated. In a way syntax and spelling errors point to human written text.

0 reply

0 recast

0 reaction

Sean Brennan pfp

Talk to @betashop.eth, Airstack worked on an XMTP filter for spam

1 reply

0 recast

0 reaction

Corbin Page pfp

@sahil is doing some great work here

1 reply

1 recast

1 reaction

Gabriel 🎩 pfp

I believe I may have been wrongly black listed. Can you please check or how do I appeal? My wallet doesn't show up at all under Airdrop 2.

1 reply

0 recast

0 reaction

simon pfp

@deployer is testing something interesting with tomatoes. If you "tip" someone a tomato, you're basically reporting a low quality or spammy cast and it reduces their AD amount. Maybe the community could help keep out bad actors with the reverse-degen version of tomatoes?

1 reply

0 recast

0 reaction

Ombre🎩💙 pfp

was in ML before cyrpto, will fosho take a look at that. I noticed an "original_cast" column that isn't in the FC database. Where does that one come from ?

1 reply

0 recast

0 reaction

iSpeakNerd 🧙‍♂️ pfp

iSpeakNerd 🧙‍♂️

@ispeaknerd.eth

@fun check this out

0 reply

0 recast

1 reaction

Mark Carey 🎩🫂 pfp

Mark Carey 🎩🫂

Some interesting metrics here. Warpcast is going to launch "power badges" soon, maybe this will help ... but the "recipe" will be secret, for good reasons. But the harder task may be identifying non-spam users who may not (yet?) be ⚡ power users.

1 reply

0 recast

1 reaction

Mateusz🎩 pfp

Maybe a measure which identifies the count of "Exactly the same messages casted"? Or something similar like measuring diffs between casts content :) Bots have a tendency to post exactly the same message many times, repeating every few minutes/hours on different channels.

0 reply

0 recast

0 reaction

Jason pfp

I work in the ML space and specifically have dabbled in NLP for 8 years. Will give it some thought

0 reply

0 recast

0 reaction

Mateusz🎩 pfp

Big step for $DEGEN community improvement! 2137 $DEGEN 🔥🔥🔥🔥

0 reply

0 recast

0 reaction

cryptonomori pfp

@cryptonomori.eth

Is it possible that distributing $DEGEN will be recognized as spam?

0 reply

0 recast

1 reaction

$neged 🔄 🎩 pfp

$neged 🔄 🎩

you can send everyone spamming to /neged we'll take them in.

0 reply

0 recast

1 reaction

yaffle 👽🎩💊 pfp

yaffle 👽🎩💊

1/ Just shared with with an ML friend of mine - his feedback was that it's a pretty simple classification algo, so something like random forest or bayes could work, however some teams at Google have been having success with LLMs using the Gemma framework.

1 reply

0 recast

0 reaction

coinbuza🎩🔵🥈🧀 pfp

coinbuza🎩🔵🥈🧀

0 reply

0 recast

0 reaction

Tyga pfp

My first thought would be to be get all the profile info of accounts, model citizens that you know for sure aren't botting and also those that for sure are. Then train your model to recognize those patterns with a score of 0-1. In betweens may be varying levels of spammy, but may have some thoughtful behavior.

1 reply

0 recast

0 reaction

Eitzi 🎩 ⌐🆇-🆇 🌎-‘ pfp

Eitzi 🎩 ⌐🆇-🆇 🌎-‘

That’s cool. What is called a „original cast“?

0 reply

0 recast

0 reaction