Dan Romero on Warpcast

Dan Romero pfp

The spam filtering dilemma Sharing how we think about it so people can better understand: 1. Most people will only use social networks if they are fun 2. Casting something and getting dozens of spammy replies is not fun 3. If you get enough spammy replies, you'll just stop using the app 4. So, aggressively filtering for spam is necessary to keep your existing users happy 5. However, this means some good new users will hit spam filters. Understatement: this is a bad user experience for those people. It would be great to not have this happen. 6. But given the choice of who to make unhappy — an existing user who has invested a ton of time and energy into the network or a new user that is most likely to churn (top tier social networks lose 50% of users!), then you have to optimize for your existing users. 7. So does that mean you don’t care about new users? No! It’s existential to have new users be able to join Farcaster and find people to connect with — without being labeled as spam. 1/2

33 replies

92 recasts

481 reactions

Dan Romero pfp

8. So here’s what we are doing to improve this so new good users aren’t mislabeled as spam: a. Continually refine the spam model by identifying incorrectly labeled people (that feeds back into the model and makes it better); good existing users on the network reporting incorrect labels is really valuable here b. Warnings to users before they are labeled spam so they have a chance to adjust their behavior c. Increase the number of signals our model uses to help filter out bots vs. people d. Working a new version of channels that give owners / moderators complete control of who appears in their channel (so a Warpcast spam label won’t affect casts in the channel) 2/2

16 replies

3 recasts

250 reactions

Pete Horne pfp

I look into the hidden replies and I think you’re doing a good job there. I don’t know how my casts go so maybe an idea is some sort of score report on my casts so I can see how to improve my engagement to help me help the quality effort. I think you’re doing a great job experimenting on delivering a quality experience.

1 reply

0 recast

3 reactions

frederick pfp

I think that the team is looking at it from all angles and trying to determine the best approach to combat a spam issue while ensuring existing users have a good experience. This is more than you would get from many other platforms.... You can't really make everything public because then the spammers would know how to get around it and the user experience would be awful.. what i've most heard from fc users is that if they reply to people that have a lot of followers - they want to know if those people not responding to their replies because: -it's not interesting -they don't see it -they run out of time -they are labeled as spam -other i've heard from those that say they stop responding bc they don't want to be spammy or labeled as spam. personally, i tell them to engage in the communities that they do get responses in and to practice patience. (acknowledging this is not always what people want to hear.) that said - i wouldn't expect to get a response from elon, zuck, etc either.

1 reply

0 recast

3 reactions

Apurv pfp

a & c were expected, b is something very interesting. Might be an additional data point for spammers to fine tune their messaging strategy basis this feedback. Guess one of those cases where false positive is more costly than false negative? (positive = labelled as spam )

2 replies

0 recast

0 reaction

bradq pfp

What's the best way to report incorrect labels? I see numerous people seemingly struggle to get help. Is it better if older, more reliable users report them to you guys and what's best way to do that...tag you or v?

0 reply

0 recast

2 reactions

Yassine Landa pfp

@yassinelanda.eth

can you guys share (privately if better) your models accuracy/recall/f1 scores? We're achieving 99.8% accuracy on https://paperswithcode.com/dataset/sms-spam-collection-data-set @v we can share notes!

0 reply

0 recast

1 reaction

Jacque(she|her).⌐◨-◨ pfp

Jacque(she|her).⌐◨-◨

It is a tough line to tell the real folx to adjust and keep them out vs the bots and farms that can run algos to adjust. I really hope that this helps but bring in that shadowban for months I know how hard it is to feel like we are not getting heard it hurts.

0 reply

0 recast

1 reaction

Push 🎩🤌🏻 pfp

Push 🎩🤌🏻

I think I have some example for point a. So how does it work? Do I just tell you here?

0 reply

0 recast

0 reaction

flick the dev 🎩📦 pfp

flick the dev 🎩📦

d. feels like a really good option to allow communities to better form their own norms

0 reply

0 recast

0 reaction

Kyle 🍖🦄🐹 pfp

Kyle 🍖🦄🐹

FWIW - the /lp channel has done a great job at promoting ham while tossing spam. The @automod plus tipping via likes and cast actions that @deployer implemented is very enjoyable!!

0 reply

0 recast

1 reaction

Anemale🎩↑🔵 pfp

Anemale🎩↑🔵

It's interesting to tackle spam, but is ‘gm' considered spam? From a cultural perspective, 'gm' is a greeting, a movement, and many people use it. I often say to provide text with content and a response appropriate to the context, but some do not do so. Will they be categorized as spam?

0 reply

0 recast

1 reaction

𝖉𝖞𝖈𝖕¹ᵍ pfp

𝖉𝖞𝖈𝖕¹ᵍ

i just want to be clear i had no issues with your spam model so you can blame your lack of sleep on every other bozo. i just wanted you to shave.

0 reply

0 recast

0 reaction

AlPlanet Ⓜ️ 🎩 🎭 pfp

AlPlanet Ⓜ️ 🎩 🎭

I love all this 🙌 A multi-sig type moderation could be a good addition maybe. Not multi-sig channel ownership, but a dynamic team (most appreciated posters) who can tag spam in the channel (open to public scrutiny). Owner would have overall say, to permaban or perma-appoint a user to the team. Casts that reach a threshold of reports are spam. No incentive, regular users cannot report. Long term; additional reputation system to find good reporters, allow them flag official project accounts with their nft contracts. The Channel Team may help with appeals 🤷 and also maybe signal the range of tolerance to spam in different channels to aid your models.

0 reply

0 recast

0 reaction

Thomas Walsh pfp

So, would any of my recent castings be considered spam?

0 reply

0 recast

0 reaction

highplains.base.eth 🎩✨🔴 pfp

highplains.base.eth 🎩✨🔴

Sounds like a great strategy. Good work.

0 reply

0 recast

0 reaction

DegenFans 🎩🔵🫂Ⓜ️ pfp

DegenFans 🎩🔵🫂Ⓜ️

Does it make sense to give new user a puppy protection or will this also be used by bots?!?

0 reply

0 recast

0 reaction

J Finn 🎩 pfp

I think c) is the most interesting space. For example for myself: - good Open Rank (a function of Followers partially) - had pb - shares music so gets some tips on /sonata - shares the odd piece of good content so earns some /degen - etc There has to be a way to create “trust signals” behaviourally in a way that scales, and I think tokens/socialfi are a part of that tbh

0 reply

0 recast

0 reaction