5 ms·

Just out of curiosity, isn't that feasible today to implement some machine learning to stop spammers? Is there any project trying to come from this angle?

lucb1e 7y ago
That's like saying "why don't you use algoritms and code". Like, sure, but what is it you're proposing? What features would you learn from and match against? (For those unfamiliar with algoritms and code as solution, it's a reference to this: https://www.reddit.com/r/ProgrammerHumor/comments/5ylndv/so_thats_how_they_did_it_its_brilliant/ https://www.reddit.com/r/ProgrammerHumor/comments/5ylndv/so_... )
- StreamBright 7y ago
  Actually we have implemented something like that for HTTP requests. Features would be: IP (first 3 octets are probably enough), posting time, length, time to solve captcha, time between clicks, country where the IP is located, post contains certain words (can be learnt from spam posts), does the post contain a link(y/n) I think I would start with these, probably looking into what other people are doing.
  - Ayesh 7y ago
    An ideal machine learning implementation would also need the context, such as the original post itself, parent comment(s), other comments in the thread, etc. It can be quite difficult than one might think. For example, now that we are talking about spam, the word "Viagra" shouldn't block my comment, even though my parent post doesn't mention the word or in a situation where nobody else mentioned it.
Ayesh 7y ago
Akismet has been doing this for some time I believe. The downside is that you have to send the form submission and other meta data (IP address, etc) to them. This can be better than the current reCaptcha v3 though.
- StreamBright 7y ago
  Good to know! I am going to check this out.
kqr 7y ago
Sure. I know a developer at Distil Networks[1] and that's exactly their core business idea. Apparently it's very hard, though, because they have positioned themselves as an alternative for very high-value targets (like airline tickets) and charge accordingly for their services. It also seems to require some amount of human intervention still. [1]: https://www.distilnetworks.com/ https://www.distilnetworks.com/
- StreamBright 7y ago
  Good to know, I guess it boils down what sort of features you can use, which might be hard to find. Based on my experience feature hunting can be a lengthy process.

lucb1e 7y ago

That's like saying "why don't you use algoritms and code". Like, sure, but what is it you're proposing? What features would you learn from and match against? (For those unfamiliar with algoritms and code as solution, it's a reference to this: https://www.reddit.com/r/ProgrammerHumor/comments/5ylndv/so_thats_how_they_did_it_its_brilliant/ https://www.reddit.com/r/ProgrammerHumor/comments/5ylndv/so_... )

StreamBright 7y ago
Actually we have implemented something like that for HTTP requests. Features would be: IP (first 3 octets are probably enough), posting time, length, time to solve captcha, time between clicks, country where the IP is located, post contains certain words (can be learnt from spam posts), does the post contain a link(y/n) I think I would start with these, probably looking into what other people are doing.
- Ayesh 7y ago
  An ideal machine learning implementation would also need the context, such as the original post itself, parent comment(s), other comments in the thread, etc. It can be quite difficult than one might think. For example, now that we are talking about spam, the word "Viagra" shouldn't block my comment, even though my parent post doesn't mention the word or in a situation where nobody else mentioned it.

Ayesh 7y ago

Akismet has been doing this for some time I believe. The downside is that you have to send the form submission and other meta data (IP address, etc) to them. This can be better than the current reCaptcha v3 though.

StreamBright 7y ago
Good to know! I am going to check this out.

kqr 7y ago

Sure. I know a developer at Distil Networks[1] and that's exactly their core business idea. Apparently it's very hard, though, because they have positioned themselves as an alternative for very high-value targets (like airline tickets) and charge accordingly for their services. It also seems to require some amount of human intervention still. [1]: https://www.distilnetworks.com/ https://www.distilnetworks.com/

StreamBright 7y ago
Good to know, I guess it boils down what sort of features you can use, which might be hard to find. Based on my experience feature hunting can be a lengthy process.