The other day, I proposed a sort of Twitter analog to PageRank that readers generously dubbed “TunkRank”. I know that some readers started looking into implementing it, so I wanted to put out an offer.
If anyone implements this measure or one that preserves its fundamental principle of representing attention scarcity, I’ll promote it prominently on this blog (e.g., a link on the front page). It has to be a web application that anyone can use, and you have to explain how the measure works. No need to share the source, though I won’t complain if you do. And I’m not asking for any rights to the work.
If you have questions, contact me or just ask them in the comments here.
The Noisy Channel 
13 replies on “The TunkRank Implementation Challenge”
[…] RSS ← The TunkRank Implementation Challenge […]
LikeLike
Here’s something related:
http://www.techcrunch.com/2009/01/18/rt-techcrunch-retweetist-discovers-most-valuable-users-and-accounts-from-twitter-retweets/
LikeLike
True, it’s related. But I have a lot of problems with retweet counts as a measure. On one hand, it’s easily gamed: I could use bots that no one follows to retweet a post. On the other hand, I think that retweets only represent a special case of propagating the message. Certainly replies should count, as should derivative messages that aren’t posted as retweets.
Maybe all of these indicators are correlated, but I’m skeptical, especially given that so many people know that retweets are being measured. I’m still sticking to TunkRank, especially since it has such a cool name!
LikeLike
Is the raw data of followers and retweet rates available? How big is the graph in terms of nodes and edges? How skewed is the following distro?
Can we really ignore how active readers are? And won’t retweet rates vary dramatically across both tweeters and tweets?
“TunkRank” is good in that it’s (a) discriminative, and (b) easy to spell and pronounce, but it’s bad in that it’s (c) non-descriptive, and (d) self-indulgent.
LikeLike
The raw data of followers is certainly available through Twitter’s API, and indeed a number of existing applications make use of it do compute their measures of Twitter authority. Retweet rates are another story–you have to determine what is a retweet heuristically by mining the messages.
My problem with using retweet rate as a measure is twofold. First, it it is too vulnerable to gaming–I could write bots to follow me and retweet my messages. Given the low number of retweets, it wouldn’t take much to radically skew the rankings. Second, retweets are too narrow a measure of influence. How about replies? Or posts that are derivative but don’t identify themselves as retweets or replies?
I’m more inclined to multiply a user’s TunkRank (yes, I do feel self-indulgent!) by his or her posting rate. Posting more frequently, all else equal, does give you more influence. If your posts annoy people, they’ll stop following you–so the system self-regulates.
LikeLike
I was tweeting about the same idea recently. I think looking at the ratio of following/followers might be a good starting point. However, in my opinion, the problem is that one can easily cheat by first following a lot of twitters to get followed and then stop following other twitters. If the TunkRank is mainly based on this following/followers ration, it might be easy to cheat to get a high score.
LikeLike
Actually, it’s not based on ratio, though I suspect it correlates to ratio. For example, I have a high ratio of followers to people I follow–which means that following me is unlikely to be motivated by the expectation of reciprocity. But I can’t game TunkRank by making fake users to follow me. Does that make things clearer?
LikeLike
I think the main thing missing in the model is the number of tweets/day a given person is doing. I mean I river of news man maybe generate a lot of good stuff but is also generating a lot of “bad stuff” to a lot of the followers.
LikeLike
Here’s my thought: in measures like retweet rate, posting more tweets per day should reduce your per-post retweet rate if most of your posts are uninteresting. I’d hope that you’d see a similar effect in TunkRank because fewer people would follow you. That said, I think you’re right that the attention cost of following someone is proportional to their posting rate, and should be treated as such by the model.
LikeLike
[…] couple of months ago, I put out a challenge to implement an influence measure for Twitter that acquire the personally gratifying (if […]
LikeLike
[…] Adams, who recently won the TunkRank implementation challenge, explains on his blog how he implemented TunkRank.com. He implemented the algorithm in Ruby using […]
LikeLike
I have made something that is very close to this but it actually does more than measure. It is a tool to measure and monetize. If you email me I would like to show it to you before I release it to the public.
LikeLike
[…] talked much about TunkRank in the past months, largely because Jason Adams, who stepped up to the TunkRank Implementation Challenge last year, has been leading the charge. Indeed, all I did, beyond lending my first syllable to its […]
LikeLike