Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Show HN: Using stylometry to find HN users with alternate accounts

stylometry.net

Post reply on HN
Print view View original

21–30 of 562 posts

Re: Show HN: Using stylometry to find HN users with alternate accounts

#21

Post by costco »

ufmace wrote: post #14

I wonder what's a reasonable threshold for "probably the same person". I've never had an alt on HN, and when I searched myself, it found 3 other users above 0.6, none of whom I've ever heard of before.

If it's >0.9 is you can almost guarantee it's an alt but I've seen certain matches at 0.6. The problem is writing styles change over time. Another idea I had was converting the scores which are just cosine similarity scores into percentiles (so 0.99 would be 99th percentile of certainty) to make them more human interpretable.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#23

Post by bonzini »

costco wrote: post #21
ufmace wrote: post #14

I wonder what's a reasonable threshold for "probably the same person". I've never had an alt on HN, and when I searched myself, it found 3 other users above 0.6, none of whom I've ever heard of before.

If it's >0.9 is you can almost guarantee it's an alt but I've seen certain matches at 0.6. The problem is writing styles change over time. Another idea I had was converting the scores which are just cosine similarity scores into percentiles (so 0.99 would be 99th percentile of certainty) to make them more human interpretable.

The people at 0.4-0.6 with me do share some interests. That's cool on its own.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#27

Post by costco »

Yeahsureok wrote: post #15

On the how to avoid section: Isn't running comments through a randomised translator a few times then back considered a countermeasure also? Also think it's probably poor form to list users as examples without their permission.

> On the how to avoid section: Isn't running comments through a randomised translator a few times then back considered a countermeasure also?

Yes.

> This may be out of line but isn't pg on here with a different username, Levenschtein distance of one that's not included? Or is that just a very motivated 13yo account who writes a lot of admin-esque comments.

What other pg account are you referring to? I want to see it so I can see what my algorithm missed.

> Also think it's probably poor form to list users as examples without their permission.

You're right. I'll remove that - I just wanted some examples especially for people on phones who don't feel like typing. Thanks for the feedback.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#28

Post by dotancohen »

ufmace wrote: post #14

I wonder what's a reasonable threshold for "probably the same person". I've never had an alt on HN, and when I searched myself, it found 3 other users above 0.6, none of whom I've ever heard of before.

Interesting. The highest non-me account is under 0.4 on my page. I do not believe that I have such a unique writing style - especially since half my posting is on mobile and therefore possibly slightly different than my desktop posts.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#29

Post by gus_massa »

It would be nice to make the names clickable.

I don't think the list of pg alternate account is accurate. I checked a few. They have many oneliners that is typical of pg, but the topics and style don't look similar.

I searched a few more and got better results. :)

I searched myself (that I know that I have no alternate accounts). I recognize a few users that are interested in similar topics, and I discuss/upvote them many times. But I didn't recognize most of the user of the list.

Re: Show HN: Using stylometry to find HN users with alternate accounts

#30

Post by throwup »

costco wrote: post #21
ufmace wrote: post #14

I wonder what's a reasonable threshold for "probably the same person". I've never had an alt on HN, and when I searched myself, it found 3 other users above 0.6, none of whom I've ever heard of before.

If it's >0.9 is you can almost guarantee it's an alt but I've seen certain matches at 0.6. The problem is writing styles change over time. Another idea I had was converting the scores which are just cosine similarity scores into percentiles (so 0.99 would be 99th percentile of certainty) to make them more human interpretable.

I make new accounts every so often and the accounts of mine that it found have a score of around 0.3. I'm not actively trying to defeat stylometry but it's possible I just have a particularly unremarkable writing style.
Post reply on HN