The HP paper concludes that most of the relationships identified in a social network are weak; the strong social network, the real friendships, are hidden within it.  Therefore, the authors argue, if you’re going to do viral marketing, you have to discover and tap into that hidden, more deeply connected network.  I think they went too far there.

They assume that only the strongest friendships can mediate viral ideas.  Why?.  I don’t think the study really addresses that question.  If you want to see the flow of influence through the network, you can’t just look at who is communicating with whom, you have to look at how people and network respond to communications.

A former intelligence guy explained it to me with a Cold War example.  If you see a pattern that when radio station X transmits, a large part of the Soviet Navy suddenly changes direction, you can guess that X is a command and control center.  That’s traffic analysis at its simplest.

In a network like Twitter, the equivalent would be to observe a correlation between person X’s posts and some measure of network activation, particularly something like “re-tweeting,” which has little cause-and-effect ambiguity.

I think of my methods of analyzing social media in three categories:

• Traffic analysis
• Social network analysis
• Linguistic/text analysis

Traffic analysis includes today’s typical web analytics and a bit more. Beyond the basics, it is the idea of measuring hidden events by analyzing visible patterns they cause. This is like figuring out how big a rock tossed into a pond is by measuring the waves it produces. A simple social example is to look for correlations between the number of people participating in a discussion and who is participating. If you find that there are certain people whose presence correlates to higher activity, you might infer that they are provoking greater participation. Or the cause might be the other way around – certain people only enter the fray when things get hot.

Social network analysis assumes, reasonably, that groups of people interacting on-line organize themselves into roles such as opinion leaders, connectors, lurkers and so forth, and seeks to identify who’s who in the network. Law enforcement and intelligence agencies have used these techniques to solve problems such as figuring out whose telephones to tap. If there are 200 people in a crime syndicate, but resources only allow you to tap 10 phones, how do you choose which 10? One answer is to find the 10 people who, as a group, talk to the greatest number of the 200. That way, odds are you’ll hear a bit of what everybody is talking about. What’s more, those 10 are likely to be highly influential.

Linguistic/text analysis aims to figure out what people are talking about and their attitudes (sentiment) toward those topics. This is the hardest kind of analysis because language is complex and ambiguous and computers are extremely stupid when it comes to language. However, just looking at how language changes over time – which words or phrases are growing in popularity – can reveal quite a bit.

I don’t use these three approaches in isolation. The language part is so challenging that I have relied on traffic and social network analysis to narrow down the interesting people to a handful, which makes linguistic and text analysis computationally simpler. That is really no different from the telephone tapping example, except that law enforcement generally still relies on humans to figure out what is really going on. Computers can recognize words and phrases increasingly well, but even at Fort Meade, I’m fairly sure they still need people to truly make sense of language.