You know Twitter spam when you see it—but wouldn’t it be nice if you didn’t have to see it?
Unfortunately, email-style filters, which analyze message contents, are of little help. Due to the rigors of 140-character communication, even legitimate tweets tend to read like Nigerian phishing scams, while the hucksters often hide their pitches in links. So Twitter simply puts the onus on users to report offending accounts.
But a fascinating recent study from Imperial College London suggests a new approach. Borrowing some tricks from computational neuroscience, coauthors Gabriela Tavares and Aldo Faisal have come up with an algorithm that can tell—with 85 percent accuracy—whether a Twitter account is home to a bot or (worse) a corporate shill instead of a regular person.
It’s all in the timing. By analyzing the timestamps on 165,000 tweets, the researchers found that these three user types—individuals, companies, and robots—have very distinct activity patterns. Think of it as temporal fingerprinting. The approach could eventually be used to create more effective filters for all kinds of social networks.
Click to see the time profiles for each user type:
Looking at just the relative frequency of tweets sent over the course of the day, characteristic patterns emerge for each account class.
But the new algorithm also parses user activity on a much deeper level.
Chart: Onformative | Source: “Scaling-Laws of Human Broadcast Communication,” PLOS ONE, July 2013
Could a spammer simply program their bots to mimic human behavior? “It would be very hard,” says Tavares, now at Caltech. “This isn’t a deterministic model—it’s not a matter of just tweeting at certain times of day or certain intervals. The classifier uses machine learning to build probability distributions; without knowing the parameters of the model, you couldn’t simulate the expected behavior.”
Stopping spam is a cause everyone can get behind. But as Faisal, a lecturer in neurotechnology at Imperial College, points out, the results also tell a bigger story about the informational richness of communications metadata—the seemingly trivial details of our connected lives. “It illustrates how even the most basic metadata can reveal a great deal about who you are,” he says. Something to think about, perhaps, in light of revelations that the NSA is busily skimming just such data from our telephone calls.
In fact, the new algorithm can not only tell whether you’re a good Twitter citizen or a spammer, it can even predict the timing of your next tweet with surprising accuracy—provided you’re human, that is. Turns out that people, as a group, are more predictable than robots. So much for our vaunted human spontaneity. Those annoying tweets that arrive with machinelike regularity each day? Those are actually your friends.
Tags: penn state football syria Jason Dufner
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.