Forecasting The Demise of Facebook
http://io9.com/researchers-predict-facebook-will-die-out-like-a-disea-1506843703
http://www.pcmag.com/article2/0,2817,2405347,00.asp
Well check this out, it would appear that academia is finally taking an interest in the application of epidemiological models to social virality. This is the first published paper I’ve seen on the subject since I began working on it many years ago. Unfortunately these researchers are still a few steps behind the state-of-the-art in their modeling techniques. I started applying the SIR and irSIR or SEIR epidemiology equations to modeling virality back when I started WildTangent in 1997. The use of the equations was invaluable to developing an extremely successful online downloadable game company that grew to be come the worlds largest online game network by 2011 via virality. At the time that I started using these equations and found them relevant I kept it a trade secret. I didn’t want my competitors to know that WildTangent had a secret formula that we were successfully applying to virally acquire traffic.
By 2009 when I was hired by Hi5.com, a social network in the Bay Area, I arrived on the job with my “secret equations” secure in the knowledge that I had a superior understanding of how social virality worked and was eager to apply my math to developing viral traffic at Hi5 which had 60 million unique monthly users. Hi5 had a huge hadoop cluster that we were able to mine for virality data. It did not take me long however to find that my math was failing to describe the viral phenomena I was seeing. The viral equations I had borrowed from the field of epidemiology were NOT correctly describing the dynamics of a social network. The problem actually mystified me for the duration of my time at Hi5. I was so immersed in the business that I never had time to try to figure out why my math wasn’t working. A few years later after selling Hi5 I finally had time to sit down with Mathematica and tackle the problem which had been nagging me for so long.
After several months of work I devised a new compartmental model that appeared to correctly model social networks. Although it was derived from the 4 compartment SEIR (Susceptible->Exposed->Infected->Recovered) epidemiological equations, I found that I had to add two additional equations to the model to produce a system of compartments that appeared to accurately model the viral data I had seen at Hi5. For readers who are mathematically inclined, I published the complete model and several articles explaining it on this blog last year under the category of “Quantum Economics”, links to those articles can be found here:
Without spouting differential equations, the problem with the SIR based models for describing social networks like Facebook was that they failed to capture some unique properties of social networks NOT found among viral diseases. The biggest flaw with epidemiological models was a phenomena I titled “re-engagement rate”. In the real-world diseases don’t become more infectious nor do infections last longer in proportion to how many other infected people are around you. The amount of time you spend sick with the flu is constant no matter how many of your friends ALSO contract the flu. This was the observation that caused my earlier models to break down. If I used them to model non-social virality of say, a hit game, they worked great, but when the content is “social content” related to friends the model fails… spectacularly. When the influence of “re-engagement rate” was introduced to the equations, the behavior of audiences could change dramatically compared to the more simple SIR based models…. Most notably a “social infection” under the right conditions could last indefinitely.
One of the most interesting insights from this work was the relationship between online social games and Facebook virality. While simple social virality tended to spread in a very classical “disease-like” way, culminating in a rapid immunity response from the population, social games had the interesting impact of inflating the networks re-engagement rate, thereby dramatically increasing the probability that somebody who had dropped out of the network would eventually return to it. It appeared that the apparently incidental success of social gaming on Facebook, may in-fact have played a critical role in Facebook’s ultimate growth and market share maintenance.
In other words… a SIR based model cannot accurately forecast Facebook’s decline. It is entirely possible for Facebook to maintain an audience long after an epidemiological model would predict was possible. It is also evident that Facebook, whether they have a mathematical model like this for their business or not, it at least intuitively aware of this and consistently make changes to their network designed to maximize their re-engagement rate.
The model I devised is called EUREKA and can be found here on my blog. To see the most current version of my model online in Mathematica follow this link: http://www.alexstjohn.com/WP/2013/02/19/modeling-game-monetization/
My suggestion would be that before anybody declares Facebook’s demise, they should plug Facebook’s viral data into my EUREKA model first. I believe it will produce a vastly more accurate result. Regardless of the current trend, the EUREKA model also shows that subtle changes to the networks properties can have a huge impact on traffic flow… in other words, regardless of Facebook’s historical traffic data, their destiny is still, to a large degree, entirely in their own hands.
*A link to the research paper that misapplies an irSIR model to Facebook resulting in lots of hysterical press on the subject.
http://arxiv.org/abs/1401.4208
*Links to previous articles on Quantum Economics and the EUREKA model for social virality
http://www.alexstjohn.com/WP/2013/02/19/modeling-game-monetization/
http://www.alexstjohn.com/WP/2013/01/29/the-tangled-webs-we-weave/
http://www.alexstjohn.com/WP/2013/01/26/eureka/
http://www.alexstjohn.com/WP/2013/01/18/not-dead-yet/
http://www.alexstjohn.com/WP/2013/01/16/too-fat-to-fly/
Great read, Alex. I too was *stunned* to learn that Princeton had predicted more or less the shutdown of Facebook. Really? Can a billion (if you fully believe the user figures) people switch off of something so ingrained in [far too many of] their lives that quickly? Well surely the answer is “no”, and I think you’ve made that point very clear.
Unlike a virus, the Facebook phenomena can’t go away that easily on its own, or even with the appearance of other social networks. By now there is far too much “background noise” in the form of daily users of that service to sustain a healthy, if not growing, user base for years to come. How many of the existing users were dragged kicking & screaming into Facebook by their friends, peers, etc. simply because it was the place to hook up? And now that they live there, for many of them it will be hard to find a compelling reason to jump ship.
I completely agree that the online games cemented it for Facebook with the general public. (Hell even I used to play Tetris and Sudoku there with my morning coffee!) It begs the question though… would a significant facelift of some sort to their existing games platform help to sustain the user base for longer than it would otherwise run? Almost like a virus mutation…
Yes, I could write a long boring Quantum Economics article on the weird symbiotic relationship between social networking and social games. What games did was give people hundreds of NEW ways to POKE friends they wanted to interact with. From my point of view POKE was the first and most primitive Facebook game from which all subsequent social games “evolved”. We did a really interesting experiment with this at Hi5. For Halloween we launched a Zombie epidemic “gift”, the rule was that you could send your Hi5 avatar to a friends site as a zombie and bite their avatar thereby “infecting” them. Anybody bitten could then send zombie gifts of their own to an unlimited number of friends ONCE before being having to either pay for it or get bitten again. As you might imagine it went HUGELY viral across out network… but unlike many other types of gifts we had released the zombie plague didn’t die off. It raged on our networks well beyond Christmas without abating AND despite being pretty much free, generated more revenue than all of the other virtual gifts in our store combined. To try to wind the epidemic down we made the gift PAID only although the bitten could still pass the virus on for free once. It just raged on and made more money. Finally we did a data run to try to figure out why people kept sending these things to the same friends back and forth over-and-over and we found that the Zombie gift had basically become a strong proxy for “poking” friends just to get some attention or social dialog going with them. It had a MUCH higher response rate from friends than other types of social messages, so people wanting social attention all used it. It was a great experiment in “re-engagement” maximization. We did a number of really interesting social experiments in the vein that I should write about some time… suffice it to say that the SEIR model DID NOT explain the never ending zombie epidemic.
You should definitely do a post on all the creative ways you re-engaged your userbase! Sounds like a fun read.
That was a lot more interesting than I thought it would be. Although I would say that you aught to expand ‘social games’ to ‘apps’ in general. Social services that require or strongly suggest integration with the social network (Facebook) outside of games I think would be as just as applicable.
With little to no experience in the field of modelling this kind of data I ask: Is the EUREKA model designed to be used before or during the life of a social game? IE is its primary purpose meant to be used to highlight problematic engagement areas in a game with live data or is it primarily used as a tool to figure out optimal areas to spend the most effort to get the best engagement from upon release of a game? Or both?! 😛
Two questions to answer here. I talk about games because I’m actually an expert on games and have studied their social behavior and analytic patterns extensively. Although you are certainly correct that many classes of social apps exhibit the same properties…. I could not claim to KNOW for sure, so I try to confine my confident proclamations to subjects on which I am “expert” 🙂
I’m embarrassed to say that it started out very practical and then I got carried away with solving the math problem and rendered the version of it on my site a little abstract for practical application. What I have been meaning to get around to writing is a version that takes standard game metrics that people track today and solves for a EUREKA “fit” for that data. Writing it is a bit of a PITA, not the most “interesting” problem to solve and it’s something any skilled mathematician could create using the equations I published, so I figured my work is done. In principal somebody who understood the model could set the parameters to fit measured data and get a predictive result. The trick would be to fit D1, D7, and D30 traffic drop-off rates to an exponential growth or decline co-efficient and plug that into EUREKA. Most companies do not have the analytics infrastructure to track things like re-engagement rate or their viral pipeline, so getting good quality data for the equations turns out to be the most difficult challenge with using them. I hate to recall how much time we spend hadoop dumpster diving at Hi5 to guestimate these kinds of metrics.
weirdly I might actually suggest using the much earlier SEIR version of the model for basic game virality metrics. It’s obviously simpler to understand and although it doesn’t model the critical interior re-engagement behavior of acquired audiences, I generally found that the SEIR model was fine for predicting the early life cycle of a viral game. The more advanced EUREKA model WAS rigged to be useful for forecasting game revenue including simulated traffic cohorts, specifically to help marketers understand when and how much to spend on traffic acquisition and how much it would return. I’m particularly proud of this work but suspect that I would have to devote a lot more effort to explaining it or simplifying it to make it usable by a wider audience of less mathematically inclined users.