Question

this is the use case I’m trying to figure this out for.

I have a list of spam subscriptions to a service and they are killing conversion rate and other usability studies.

The emails inserted look like the following:

[e-mail]

[e-mail]

[e-mail]

roger[…]_surname[…]@hotmail.com

---

What would be your suggestions on spotting these entries by using an automated script? It feels a little more complicated than it actually looks.

Help would be very much appreciated!

Answer 1

I don’t think you can easily check for this. It’s not likely to be a simple string matching problem that you can throw a regular expression at because I would guess that your use of the name ‘Roger’ was just an example, and that any number of names can appear in that position. You could also run one of the regular expressions supplied by the other posters, parameterising it with every permutation of obvious first name and last name. This will probably take somewhere between “too long” and “forever”, and will flag up plenty of false positives.

Another approach, which works with the pattern you posted above, would be to take the last 4 letters of the username, and compare them against something. Spotting characters that are random as opposed to arranged sensibly (given a specific language) can be done by training a Markov Chain on legitimate text which can then allow you to calculate the probability of a given 4 letters appearing in that order in that language. For random letters, this probability will typically come in far lower than for a legitimate name (although if there are special characters or digits in there, all bets are off).

Another way might be to use a Bayesian filter (eg. something like Reverend in Python, though there are others) trained on the last 4 letters of legitimate email addresses. This would probably spot 95% of the ones which were just random, providing you made the data usable. eg. Submitting not just the 4 letters but each of the 2-letter and 3-letter substrings inside it, to capture the context of each letter. I don’t think this would work as well as the Markov-style method though.

Whatever check you do, you can cut false positives by only submitting certain email addresses for it (eg. only those at webmail addresses, which contain an underscore, with at least 3 characters before the underscore and 5 characters after it.)

But ultimately, you can never know whether it’s a spam address or a real one for sure until it gets used for one purpose or the other. So if possible I’d suggest giving up on trying to analyse the content and fix the problem somewhere else. In what way are they killing conversion rate? If you’re counting these dummy accounts in some sort of metric, you’d be best off adding a verification stage first and only caring about metrics for accounts that pass verification. Some people really do have addresses like [e-mail], after all.