Better Popularity Metrics For Twitter
Imagine two Twitter accounts. One has 100 followers, the other 200. Which account is more popular? And yes, this is a trick question.
At first it might seem like the answer is braindead-simple – users with lots followers are obviously more popular (on Twitter) than those who have just a small bunch of people looking at their updates. Indeed, this is the approach taken by the recently established Twitter directory WeFollow – users and tags are sorted by the number of followers, and so far it seems to work out okay. However, as you can probably guess, I’m about to say that this simplistic approach is wrong and suggest a more “advanced” way to measure Twitter popularity 🙂
When we determine popularity based on the follower count we treat each follower as a single vote, and in good democratic fashion we consider all votes equal. But the voting analogy doesn’t really hold. The real value that’s being passed around in the Twitter-verse isn’t votes – it’s attention. If you follow someone then you’re going to devote a portion of your time and attention to read their updates. And attention is a scarce resource. The more users you follow the less of your attention each of their updates will get.
We need to take this into account when counting the “votes”. Fortunately, somebody has already worked out the math for us :
Agalmos was first conceived for the Twitter platform, where people can “follow” other people’s updates, and the number of users one is following is assumed to be inversely proportional to the attention each feed is receiving. Then we use information entropy. Look up Wikipedia.
Behold the Twitter/RSS/”subscribing model” formula to determine agalmos:
Source : “Agalmos – an information measure for agalmic economics” (Edit: Source page is down.)
I’ve written a little script that calculates the “attention value” of a Twitter account using this function. Due to API limitations it only analyzes the first 50 followers and then uses the average value to estimate the total, but I think it works well enough as a proof-of-concept. You can try it out yourself here.
To put the numbers it reports in perspective, here’s what the script says about my two accounts :
- 86 followers – mostly normal users
- 5.93 follower value
- 193 followers – mostly spammers and people with auto-follow scripts
- 7.47 follower value
It’s not perfect, but judging from this singular, anecdotal example, I think it’s clear that an attention-based metric works better than directly comparing the number of followers 🙂Related posts :
Great article. Check out http://www.TweetTop.com for who to follow on twitter in many different categories. Similar to wefollow but a mix of UGC and editorial controlled (as opposed to wefollow which lets users self classify).
I’d like to incorporate some of your thoughts into tweettop.com for a better ranking algorithm. This could, theoretically remove our need for manual editorial control.
The way TweetTop works reminds me more of AllTop not WeFollow.
Having the algorithm implemented in a real site would definitely be interesting.
This is a good idea but it has a pretty gigantic hole in its logic: it assumes an even distribution of attention across all followed accounts. In fact such distribution is wildly uneven and probably conforms to a power law. If I follow two people, and one of them tweets every month while the other tweets every minute, it’s pretty obvious that one of these people gets much more of my attention than the other. Obviously a contrived example but in real life people like Scoble get unfollowed because of excessive tweet frequency.
Probably some amount of data mining would be necessary to quantify that dimension. You’d need a reasonable picture of per-account frequency and obviously those frequencies show varying degrees of stability (simple real-life example, one person I follow tweeted way more than usual over the holidays because they were stuck in an airport and bored).
That’s a good point, and it illustrates that attention metrics can get hairy fast. For example, suppose I follow someone who lives in a different time-zone. Even if they tweet a lot, I might miss or ignore most of their tweets because I’m asleep when they arrive.