Earlier today I received a report from Mr. Tweet with a list of recommended people I should be following on Twitter. I was waiting for that report since last week when I first knew about the project. I was really looking forward for the results, because they would help me to build effective relationships on Twitter and was in fact curious on how they would do that.
When I started browsing the report, I must admit I got a little frustrated because, despite the fact I had effectively found cool interesting people to follow, I also found a long list of other people not interesting at all for me (some of them I already knew them and had explicitly decided not to follow them, some of them were complete strangers with no interest match with me). My complaint had an immediate reply from @earlkman that inspired this post (and a direct message from MrTweet trying go get to know what my suggestion to their service was).
I sent a direct message back to MrTweet guys trying to make my point in 140 chars, but I felt a more detailed explanation was needed, and that's how I got a great excuse to blog again. The point is that MrTweet service seems to be building a list of recommended people based on behavioral approaches (who do I follow ? who do my followers follow ? who are the people most of my contacts are following and I am not, etc). That's not a bad approach, in the end it was able to provide me with relevant information, but it also suggested many items (people) I immediately discovered I was not interested in.
What was wrong ? Nothing is wrong, it's just incomplete. In my opinion, this kind of recommendations are lacking a more semantic approach. . The key to success in recommending interesting people to follow is not necessarily who they follow or who are they followed by, it's the content of what they say in Twitter. The key is the type of issues they talk about, the opinions they express, the links to other sources they provide , that's relevant ! (or at least that would work as a perfect second-pass-filter to the behavioral and more traditional approach). Applying a second filtering based on concepts and semantics would improve dramatically recommendations.
So ... who's the winner ? I think Behavioral and Semantic are no rivals at all ... they are more like brothers ... and the key to success is making them work together. Fortunately, my colleagues and partners at Popego know this better than me ;)
