4 ms·
Why can I identify spam at a glance, with greater than 90% accuracy, from the subject line alone, and to Google it's an unsolved problem? Google reads and inde
by SoftTalker 4y ago
Why can I identify spam at a glance, with greater than 90% accuracy, from the subject line alone, and to Google it's an unsolved problem?
Google reads and indexes every email that they deliver to a gmail inbox. They probably have the worlds largest and highest paid staff of ML experts, and they resort to auto-flagging based on IP address?
Fix it.
- Avamander 4y ago> to Google it's an unsolved problem? It's an unsolved problem for anyone. > Why can I identify spam at a glance, with greater than 90% accuracy I don't think you understand how terrible even 99.998% accuracy is at their scale. > They probably have the worlds largest and highest paid staff of ML experts, and they resort to auto-flagging based on IP address? You think they aren't using them for this? IP addresses very likely go into their model as well, when determining something is spam. > Fix it. They are and adversaries are breaking them again.
- kingjojo23 4y ago> I don't think you understand how terrible even 99.998% accuracy is at their scale. What does scale have to do with it, am I missing something? Wouldn’t a spam detection rate of 99.999% mean that only every 100000th spam mail would get through on an account basis, how’s that terrible?
- Avamander 4y ago> What does scale have to do with it, am I missing something? The person I replied to said they can identify spam with 90% accuracy with a glance, my point was that it's abysmal as accuracy. Even orders of magnitude better it'd still be bad. > Wouldn’t a spam detection rate of 99.999% mean that only every 100000th spam mail would get through on an account basis, how’s that terrible? I didn't specify if it's per-account or a false positive rate, it's also actually not relevant to the point I'm trying to make. When you receive billions of letters a month small errors start to matter and they're doing a better job than the person I replied to thinks they do.
- denton-scratch 4y agoYou can identify the spam because it's your inbox; you know what your goodmail is supposed to look like. For that reason you can use locally-trained naive Bayes on the client, and get good recognition (you don't need ML). Honestly, the best filter in my experience is Spamhaus Zen; it spots most spam with very few FPs. It's an IP blocklist.
- Avamander 4y agoIf you use something like rspamd, then doing ML is not terribly hard. Then it becomes a problem of providing a proper teaching set. Combining RBL's (be it URI, IP or domain WL's or BL's) will help with that. In the end it's still a hard problem, so takes a bunch of work.