Using nothing more than their public twitter relationships, is it possible to predict whether a US Congressperson is a Republican or a Democrat? The answer seems to be a guarded “yes” — our tools predict correctly 40/46 times (or around 87% of the cases.)

This post follows on from a post earlier today in which I asked, “can we calculate party affiliation?” The data set in the earlier post was gathered from the 16 members of the UK parliament who are on Twitter and the relationships between them.

Tweetcongress maintains a list of US congresspeople on Twitter. Today (February 13, 2009) there are 76 congresspeople on the service, but when I collected my data set of “who follows who” on February 3, 2009 there were only 65. Of these 65, fully 19 (29%) lived a life of noble isolation with regards the network — none of their peers linked to them, and they in turn linked to none of their peers. Removing these Miss Havishams from the data set leaves me with 46 twittering congresspeople who form a network.

Now as both social network analysis and Aesop would have it, “a man is known by the company he keeps.” What I mean by this is that given the partisan nature of politics, we should expect that Democrats will link to other Democrat twitterers more often than they link to Republican twitterers and vice versa. So that’s what NetDraw[1] , the software I’m using for most of this stuff, looks for, or more accurately:

To identify factions, NetDraw software iteratively searches for a distribution of nodes among a selected number of factions to minimise the number of connections between factions and to maximize the number of connections within factions.

Whatever. So I let NetDraw loose on the data, and here’s what it did.

I coloured the nodes red for Republican and blue for Democrats[2], labeled the nodes by party (for the sake of clarity, and for the hard-of-thinking, that’s “R” for Republican and “D” for Democrat) then counted all the nodes where label said one thing but colour another. There were six of these nodes; so NetDraw got the answer right 40⁄46 of the time (just about 87%.) This is less than the astonishing 93.75% accuracy we got with the Westminster twittering members of parliament in the previous post. Nevertheless I think we can safely say that it’s not a particularly integrated (or bipartisan) network if we can predict party affiliation with quite such success.

Here’s exactly the same map with the errant sheep re-labeled with their proper names so it’ll be easier to refer to them (if it helps, you can click on the image to view or download a larger version.)

You’ll see, I hope, that NetDraw has made a pretty good fist of the job. Where it has gone wrong on the whole is where the data clearly suggests something else. So Rep. Jared Polis for instance follows (and is followed by) no Democrat peers. Rep. Nancy Pelosi (D) and Sen. Richard Durbin (D) follow each other, but since Pelosi is followed by several Republicans and none of her other Democrat peers you can see why the algorithm has made the incorrect guess that the two of them are Republicans. Long-serving member Neil Abercrombie, as discussed in a previous post on US Congress Twitter folk, forms a bit of a bridge between the two parties, so despite his membership of the Congressional Progressive Caucus and liberal voting record, from the Twitter network point of view, his affiliation is somewhat ambiguous.

Sen. McCain follows none of his peers, and appears to inherit his incorrect attribution from Sen. Susan Collins. For the life of me, I can’t work out what makes it think that Sen. Susan Collins is a Democrat. She really isn’t, you know.

Note 1: NetDraw is a free program written by Steve Borgatti from the University of Kentucky. If you’re interested in playing around with this stuff, you’ll need to get yourself a copy.

Note 2: Actually, that’s not true. Despite a friend sharing the simple mnemonic that “‘Republicans’ and ‘red’ begin with the same letter,” I just can’t get it out of my English head that the Republicans should be blue and the Democrats red. As a result I waste precious minutes re-colouring these maps in Illustrator. It is worth pointing out that I also have problems with “left” and “right” on occasion — preferring instead the binary opposition “left” and “No! no! The other left, for God’s sake!”

This is a follow-up post to Why doesn’t the Tory MP have Twitter friends? — a report on some early research into the interrelationships between the few Westminster MPs who are on Twitter.

According to Tweetminster, the number of UK MPs on Twitter has doubled since this time last month. Where there were eight Twittering MPs, there are now sixteen. Here’s the map that shows who follows whom (the labels may be too small to read — if you want to see a larger image, click on the map.

I’ve coloured each node to show party affiliation; for those of you who are unfamiliar with British politics, Labour (our left-of-centre party) shows up in red, Conservatives (our right-of-centre party) in blue, and Liberal Democrats (what it says on the tin) in yellow.

The size of each node represents the individual’s “betweenness centrality” — a network analysis term that helps us place a value on individuals within a network. To give you a sense of what it means, the higher the betweenness centrality of an individual, the greater the impact when you take them out of the network. For those of you who work in large companies, it may be worth noting that senior management’s personal assistants generally have very high betweenness — something that is mostly remarked upon when they go on holiday (simultaneous translation: “take a vacation”.)

So far so good. By now, regular readers will probably be kissing their teeth and thinking “so what?” I’ve done a lot of these Twitter maps in the past and the novelty must be wearing off on you by now.

So here’s the thing. There are a few network analysis techniques that let one identify cliques and factions. What we’ve got here is a small set where we already know what people’s affiliations should be. How interesting, I thought, would it be to see how well the calculated result fits the real world data? Here’s what I found:
Continue reading →

A couple of days ago I did a little more analysis on Republican and Democratic Congresspeople on Twitter. Towards the end of the post, I realized that the unduplicated reach pareto chart that I’d built would only make sense if the US were a one-party state (or to be fair, if both parties had a single issue that they were united in wanting to promote.)

So — wanting to make this a little more representative — I went back and produced two charts; one showing Republican unduplicated reach (which follows a typical 80:20 distribution)…

Continue reading →

Three weeks ago (and at the prompting of my colleague Eddie Garrett who heads up Porter Novelli DC’s digital team) I mapped out the interconnections between US Congress Tweeters. We’d been working on a Twitter crawler and it seemed like a good opportunity to test things out on a new data set.

This is a follow-up post. Once again it was prompted by a third party: Christie Findlay at Politics Magazine asked whether it would be OK to print a copy of one of the maps in their March edition. I’ve heard that three weeks are a long time in politics, so I thought I’d better run the crawl again just in case. Also I’ve got a new crawler that uses the proper Twitter API (I can see some of your eyes glazing over you know. Just skip ahead when that happens.) I’d tried it out on the Porter Novelli data set, but welcomed a chance to try it on something more meaty.

So yesterday morning before work I ran the crawl. I use the excellent Tweet Congress as my source of information about which congress people are on Twitter.
Continue reading →

Last weekend I posted a chart of Porter Novelli Twitter folk and their followers. If you read it, you’ll recall that I was dissatisfied by what it implied about the collective reach of Porter Novelli twitterers.
Well, thanks to a long-ish train journey to Bolton and back, I was able to fudge a little perl script together to look through the data to find and remove everything other than the first instance of a follower. Let’s make that a little clearer. Let’s say that we’re looking at three Twitter people, Alice, Bob, and Carol. The first thing to do is to see who follows them:

Now we need to rank them in order of “who has the most followers” (also known as “popularity” as it happens). Here I’ve done that from left to right. Bob has the most followers and Carol the fewest.

And finally we go through from left to right removing all followers who have already shown up on someone else’s list.

Bob, being at the top of the list gets to keep all his followers which may seem unfair. But it’s not unfair if the question we’re trying to answer is “how do I reach as many people as possible by speaking to as few people as possible?” That is, I’m looking for reach (marketing people often express themselves in terms of “reach” — or the number of people who are exposed to a message — and “frequency” — or the number of times the average person is exposed to that message.)

Looking at the example above, we can see that Alice really delivers an incremental benefit of two new people, and Carol only reaches one new person. That gives us a much better idea of how valuable the most popular person (Bob) really is.

Applying this to the Porter Novelli data set

Clearly it would be extraordinarily boring to perform the process described above for the 205 people in the Porter Novelli data set that I want to analyse. But the analysis script that I wrote (with plenty of help from the perl monks) goes through exactly these steps. It’s a pretty straightforward job, ranking and deduping. Here’s what we get.

This makes much more sense than the last run. According to the Pareto principle, roughly 80% of the effects should come from 20% of the causes. Here we see that 20% of the Porter Novelli Twitter users (marked in black) account for slightly more than 80% of the reach (marked in red.) It’s pretty much a text-book example. Things are as they should be, I suppose.

More to the point, we can now assign appropriate value to coverage at the head of the graph. This is of great value when thinking about our media planning and engagement

By the way — if you’d like a copy of either the Twitter follower API query engine (it’s a well-behaved command-line thing that was developed by the excellent Joachim Larsen) or the slightly shonky perl script that I wrote on the train, you have only to ask: I’ll be pleased to share. Send me a tweet at @mediaczar and I’ll send you the scripts.

Yesterday I did a little work with the TwitterCounter API. Today I’ve gone a little further and (purely as an experiment) ranked a list of Twitter people in Porter Novelli by the number of their followers.

What happens if we chart this? Here’s a kind of Pareto chart showing users ranked in order of followers and the total reach that we get at each stage.

If you’ve seen this kind of thing before, it looks wrong, doesn’t it? That red curve should be steeper at the beginning and have longer flatter asymptote. If you’ve ever heard of the 80/20 rule this is one of the graphs that describes it. Normally the head of the graph (the first 20% of the x-axis) controls around 80% of the value while the tail (the remaining 80% of the x-axis) controls around 20% of the value. If you’ve ever heard about the long tail, it’s this tail that Chris Anderson et al. are talking about.

What’s wrong with the data?

It’s not so much the data as what I’ve not done with it. There must be many, many duplicated connections here. So now I need to write something that will go through the followers of all the Porter Novelli Twitter usernames in ranked order, and only count unique (or unduplicated) followers.

I’m hoping that when I re-do the chart, it will look something more like this:

Three days after my last map, and after lots of internal nudging from our CMO Marian Salzman, her two helpers Tikva Morowati and Zeenat Duberia and local activists like Juriaan Vergouw, Burçu Kaptan, and Umut Ersoy, the map of Porter Novelli people on Twitter looks very different. (You can click on any of the maps in this post to go to their Flickr page where you can choose to see them at larger sizes.)
Continue reading →

Marian Salzman (our Global CMO here at Porter Novelli) has had the inspired idea of getting people in the agency to tweet about the most exciting story this week (probably) — the inauguration of Barack Obama

You can see the results of the experiment on her blog.

I’m all for this, of course, for several reasons:

1. It gets new people onto Twitter
2. It helps us create a stronger network among Porter Novelli twitterers
3. It means I can track who at the agency is on Twitter

Continue reading →

This is a map of the current US congressmen and women who are currently on Twitter (you can click it to see a bigger map where you can read the names.) The direction of the arrows show who follows whom, and the size of the blobs indicates how “popular” a given congressperson is among their twittering peers (where “popular” means something like “is followed by many of their peers.”) Colours indicate party affiliation (for those of you who — like me — don’t live in the ‘States and who — like me — need reminding from time to time, the Democrats are the blue dots.)

Network of US Congress twitterers showing citation frequency. Click for bigger.

A cursory glance at this map shows a few things:
Continue reading →

On November 10th, Stephen Davies collected together a list of “UK PR people on Twitter” According to PostRank, this (and his earlier post, “UK Journalists on Twitter“) are the most popular posts on his blog.

Then a couple of days later, Stephen Waddington pushed that list through TwitterGrader to come up with his list of “Top 50 UK PR people by Twitter influence”

A couple of weeks ago, I was looking for a seed list with which I could test our “whitelist” and “canonify exception” rules on Rufus (the network analysis tool that Porter Novelli has been working on for the past six months.) This isn’t the right place to go into it, but to put it simply, the whitelist restricts the search to domains that are on the list (like a guest list), and the canonify exception list stops Rufus from chopping the subdomains or directories off the list (without this, a site like sethgodin.typepad.com would just show up as typepad.com or en.wikipedia.org/wiki/Social_network_analysis would show up as wikipedia.org. Rufus, by the way, is named after the George Carlin character in Bill & Ted’s Excellent Adventure.

My colleague, Tim Hoang used to work with Stephen W., so he sent him the image. Wadds then posted “the map on his blog“. My flickr page has never had so much activity.

Here’s the original graph:

Lots of people started drawing conclusions about the nature of PR, or the nature of Twitter from the graphs. There was lots of interesting speculation. Some people thought that this demonstrated how introverted the twitter crowd is. Others thought that it showed how introverted the PR/Social media crowd is. Others seemed to think that it didn’t matter.
Continue reading →