A Quick Analysis of Twitter Conversations on Osama Bin Laden’s Demise

I was traveling yesterday and felt like I missed a lot of the conversation about the demise of Osama Bin Laden, so wanted to try and catchup a bit today.  We set up a Twitter collection yesterday to capture conversations about Bin Laden in the wake of the news.  The volumes, as noted elsewhere, were pretty outstanding.  On one connection we saw 15,000 tweets per minute and another 5,000 tweets per minute, which both paled to Twitter’s new sustained tweet record – 3,000 per second.

From our data pull we grabbed all the tweets coming from mobile devices and plotted them against time.  We also ran the tweets through a sentiment analysis engine to see how people wwere reacting to the news in different locations across the globe.  The resulting map is below:

Since these were just tweets from mobile devices the sample size is a bit low, but we have high level of confidence of the location of the user.  Generally speaking you should not be representing a tweet with an exact point location unless it is coming from a GPS enabled mobile device.  Otherwise the location can be incredibly misleading.  While the map tells an interesting story it is a bit difficult to see a pattern in the data.  So, I did a quick aggregation of the data to country level, and calculated average sentiment for each.  Since we are aggregating the data to countries we can also include the geo-ip data at the city level.  We know the accuracy of the georeferencing is to the country level, so we can show trending without presenting misleading data.  Finally, to give a perspective on sample size, I added graduated symbols showing the number of tweets coming from each country.  The result of the map is below:

The countries in blue are locations where there was positive sentiment around the discussion of Osama Bin Laden and the countries in red is where the sentiment was negative.  In the case of this discussion happening around Bin Laden’s death the analysis is tricky because a tweet could be positive about the death of Bin Laden but negative about Bin Laden as a person or the other way around.  So, this quick and crude analysis should be taken with a very large grain of salt and referencing back to the raw tweets is necessary.  that said it is iintersting to see the positive support for Bin Laden in Saudi Arabia, Pakistan, Malaysia, and Kenya all locations that have been linked to Al Queda in the past.  Guatemala and France are interesting outliers.

Another challenge with this particular analysis, as mentioned previously, is the relatively low data volume.  For this pull we used Twitter’s “garden hose” which provides a 10% sample of all tweets.  Over all we pulled 1,781,969 tweets of which 6,458 came from mobile devices and 22,516 came from geo-ip.  That means about 1.6% of the tweets had meaningful location attributes.  If we also consider that the “garden hose” only gave us a 10% sample of all tweets that mean we are looking at .16% of all tweets.  At first blush this is a discouragingly low number, but if we put it in the context of a poll and calculate a margin of error with a sample size of 28,974 we are 95% confident of being +/- .58% accurate.  Using the tweets as a poll works quite using it as a census of the population, on the other had, is very challenging.  Doing geographic analysis of trends and sentiment with Twitter can be incredibly powerful, but needs to be caveatted appropriately as well as have bounds set on how meaningful the sample sizes are.  The potential, when harnessed correctly, gives us one of the first examples of realizing a true human sensor web.