Debugging BlogPulse

I love to use BlogPulse. I always get a kick out of seeing trends like this:

Or this:

These graphs often have a straightforward story behind them, allowing for a reasonable comparison between mentions of different words. Note, of course, that the graphs show the percentage of blog posts that contain a term, giving a normalized view.

But, what could explain something like this:

Here we see a term 'movie' which appears to have some sort of seasonal trend, dipping in autumn and rising again in the winter. However, the term 'guitar' appears to have a very odd shape, with a dramatic and sharp increase in the winter. Looking at this term on its own, we see:

There is a reason for this. If we look at links to weblogs published on MySpace, we see a matching pattern.

The reason that there are changes in the number of blog posts which link to MySpace blogs is that BlogPulse (Nielsen Online) is adjusting its crawling strategy over time (the above suggests that there was an increase in July, a decrease in September and another increase in October).

So, while I continue to believe that BlogPulse, and the trending tool in particular, are very useful, one has to be careful (and informed) regarding the base data that these analytics are built on.

 

Comments

[Johan]

The game "Guitar Hero III: Legends of Rock" was released on October 28. That's IMHO the reason for the increase of "guitar" in the winter.

[Gclectic]

I was also going to jump in and mention the October release of Guitar Hero 3, which seems to account for most, if not all, of that peak -- but Johan beat me to it. Add in the influence of "Rock Band", and you might be able to account for one or two extra bumps in the graph.

[Matthew Hurst]

A search for 'guitar -hero' shows the same pattern. The implication I was making in the post was the posts on MySpace are more likely to mention 'guitar'. Have a look at this graph which shows 'guiater -hero' and '"guitar hero"': http://blogpulse.com/trend?query1=guitar+-%22guitar+hero%22&label1=&query2=%22guitar+hero%22&label2=&query3=&label3=&days=180&x=12&y=14

[the constant skeptic]

I believe that there are always going to be anomalies in the data as long as the collection strategies of blogpulse remain limited.