Going viral is going to be harder than I thought.

I recently advanced an idea regarding the Black Queen Hypothesis for advertising – basically letting others do the heavy lifting to promote a campaign.  I still think the idea is valid, but recent research suggests that what goes viral in social media (like Twitter) may be completely random.  At least that’s the theory being put forth by researchers at Indiana University.

In an article published in Scientific Reports (DOI: 10.1038/srep00335) the authors built a computer simulation designed to mimic Twitter.  In the simulation, each Tweet or message was assigned the same value and retweets were performed at random.  Despite this, some tweets became incredibly popular and were persistently reposted, while others were quickly forgotten.

By analysing 120 million retweets – repostings of users’ messages on Twitter – by 12.5 million users of the social network, researchers at Indiana University, Bloomington, learned the mechanisms by which memes compete for user interest, and how information spreads.  The reason for this, says team member Filippo Menczer, is that the simulated users had a limited attention span and could only view a portion of the total number of tweets – as is the case in the real world. Tweets selected for retweeting would be more likely to be seen by a user and re-posted. After a few iterations, a tweet becomes significantly more prevalent than those not retweeted. Many users see the message and retweet it further.

“When a meme starts to get popular it displaces other memes; you start to pay attention to the popular meme and don’t pay attention to other things because you have only so much attention,” Menczer says. “It’s similar to when a big news story breaks, you don’t hear about other things that happened on that day.”

According to the authors, the wide adoption of social media has increased the competition among ideas for our finite attention. They employ a parsimonious agent-based model to study whether such a competition may affect the popularity of different memes, the diversity of information we are exposed to, and the fading of our collective interests for specific topics. Agents share messages on a social network but can only pay attention to a portion of the information they receive. In the emerging dynamics of information diffusion, a few memes go viral while most do not. The predictions of their model are consistent with empirical data from Twitter, a popular microblogging platform. Surprisingly, they can explain the massive heterogeneity in the popularity and persistence of memes as deriving from a combination of the competition for limited attention and the structure of the social network, without the need to assume different intrinsic values among ideas.

If you work in social media at all, I highly suggest that you click here to read the entire article.  A review of responses to the paper indicated that their position is a bit controversial and certainly not everyone agrees, but the elements put forth, especially in the absence of emotional factors, may be highly enlightening.  As an agency, we look for ways to go viral – it’s not quite the Holy Grail, but it’s great for awareness and usage.  The ROI on going viral is tremendous – and usually results in a much greater reach than what could be achieved by placed media.  The challenge, as with all advertising, is if the effort is reaching the right people – but get popular enough and it won’t matter.  The cost per impression is so small that it’s going to hit enough of the right people to make it worthwhile.

The authors present several elements to their research, but a couple of items jumped out at me.  The first is the impact of limited attention.  They first explore the competition among memes. In particular, they test the hypothesis that the attention of a user is somewhat independent from the overall diversity of information discussed in a given period. They quantify the breadth of attention of a user through Shannon entropy S = −Σ_i f(i) log f(i) where f(i) is the proportion of tweets generated by the user about meme i. Given a user who has posted n messages, her entropy can be as small as 0, if all of her posts are about the same meme; or as large as log n if she has posted a message about each of n different memes. They can measure the diversity of the information available in the system analogously, defining f(i) as the proportion of tweets about meme i across all users. Note that these entropy-based measures are subject to the limits of our operational definition of a meme; finer or coarser definitions would yield different values.  The key observation here is that a user’s breadth of attention remains essentially constant irrespective of system diversity. This is a clear indication that the diversity of memes to which a user can pay attention is bound. With the continuous injection of new memes, this indirectly suggests that memes survive at the expense of others.

The other is the impact of user interest.  It has been suggested that topical interests affect user behavior in social media.  This is a potentially important ingredient in a model of meme diffusion, as an interesting meme may have a competitive advantage. Therefore they explore whether user interests, as inferred from past behavior, are predictive of future behavior.

The authors consider every user in the dataset and any retweets they produce. When a user u emits a new retweet, they define her interests I_u as the set of all memes about which she has tweeted up to that moment. They also collect the set M₀ of memes associated with the new retweet. The n most recent posts across all users prior to the new retweet are considered as a set of potential candidates that might have been retweeted, but were not. The corresponding sets of memes M₁, M₂, …, M_n are recorded (n = 10). We compute the similarity sim(M₀, I_u), sim(M₁, I_u), …, sim(M_n, I_u) between the user interests and the actual and candidate posts, and recover the conditional probability P(retweet(u, M)|sim(M, I_u)) that u retweets a post with memes M given the similarity between the memes and her user interests. We turn to the Maximum Information Path similarity measure^31, 32 that considers shared memes but discounts the more common ones:

where x is a meme and f(x) the proportion of messages about x.