How does Stack Overflow determine that a question was upvoted by the questioner's sock puppet?

Question

How does Stack Overflow determine sock puppet accounts that give "fake" upvotes to questions?

I am specifically interested in the statistical approach Stack Overflow takes (IP addresses or user-agents need not be part of the answer). It's okay if propriety details are not revealed. But we can still benefit from knowing the general approach Stack Overflow takes.

My gut feel is that Stack Overflow uses a record of each user's voting decisions (e.g. last 50 votes cast by the user). They use this historical data to calculate whether a vote currently being cast can be flagged as a socket-puppet or not (i.e. it's above some Bayesian probability threshold). And when enough sock-puppet votes materialize, Stack Overflow takes action against the profiles involved.

Could an expert shed light on how Stack Overflow does it?

Answer 1

Some of this stuff is done automatically, and some of it is done manually. There is a system that checks for voting fraud by looking at voting patterns and reverses suspicious sets of votes. (The details are kept secret to stop people from trying to get around the limitations.) However, that's all it does - it won't suspend users automatically. That's where moderators step in. We can't see individual votes, but we can see aggregate data and totals - i.e. how many votes were cast by User X on User Y's posts. We can also see if a user's votes were reverted by the system.

If we notice some suspicious activity somehow, we'll try to figure out if there's malicious intent behind voting, and we'll take appropriate action, which usually involves suspending the user and deleting sockpuppet accounts. We can use various tools that make it easier to figure out if two accounts were run by the same person - we're not just guessing based on vote totals.

TL;DR: Reverting serial voting is done automatically. Suspending/deleting accounts is done manually.