Facebook connection targeting: which friends are the best?

My friends aren’t all the same.   And neither are yours, I bet.   Same thing when it comes to social graph marketing:  some friends are better than others at inspiring friends to action.    The question is- which ones?

Right now, there’s a lot of interest in methods for effectively identifying ‘influencers’ based on identifiable characteristics, such as their connectivity in a social network.   The question isn’t a new one  – my gut feeling is that it is as old as human communal life.    However, what is new is that online social networks provide the means, motive and opportunity for understanding these effects at scale, and putting them to work.    Hence the interest.

I’ve written recently about the findings of a recent study by Wei, Jang, Adamic, de Araújo and Rehki that looked at social game invitations and outcomes in two popular Facebook games from LOLApps, Diva Life and Yazuka Lords.  The paper is called Diffusion dynamics of games on online social networks, and it was presented last year at the Usenix conference Workshop on Online Social Networks WOSN 10.   (If you want to read it directly,  there’s a .pdf  of it here. ).    The headline result was that players recruited via friends were more engaged, and played longer.

But there is a lot of tantalising information in the article about what type of invitation behaviour is the most successful.   Here are some of the key results from this part of the work:

• people who have more friends invite more friends
• the success rate for invitations decreases strongly as more invitations are sent
• the success rate for invitations decreases strongly as the number of invitations sent in one batch rises, controlling for the total number of successful invitations made

Very clearly, there is a story to be told here about more selective invitations being more effective, when measured per invitation.    But there is more than one story that can be told about what this means in practice, for designers.

There are also lots of potentially interesting stories that didn’t make it into the paper.   For one thing – it’s not clear whether the success rate for invitations varies with the proportion of friends invited, as well as with their absolute number.    It seems likely that there would be a difference in outcomes between someone with 500 friends inviting 5 friends, someone with 50 friends inviting 5 friends, and someone with 5 friends inviting 5 friends.     Similarly, someone with 500 friends who invites 50 friends is quite a different type of fish from the person with 50 friends who invites 50 friends.     Why does it matter?   It matters because you’d expect the success rate to be different for these cases, and success rate, rather than the total amount of success, is increasingly important (for reasons I will explain later).

Before we mope too much about what we don’t know as a result of this work, here is something we do know:

• 10% of users are responsible for 50% of successful invitations

Wow.   That’s pretty much all I can say about it.  Except maybe awesome (which, it should be noted, I said with an entirely straight face 😐 ).

We’re not told in the article whether these top performing ‘salespeople’ are more accurate, or more prolific – or both.  We don’t know what their success rate is.  And that’s a very important question.

Are the top 10% responsible for their fair share of unsuccessful invitations, or are they successful because they are so prolific it doesn’t matter if they are even averagely effective?    My guess – and it is purely that, a guess – is that numbers game at the moment is such that super-promiscuous inviters, who have lots of friends, and invite them indiscriminately, and repeatedly, give the highest absolute return in terms of new eyes on screens, despite being the least efficient.    Kind of like the person we have all met who is successful in making conquests  because he (or she) really doesn’t care about failures, only about successes.

Targeting all of a person’s friends will give the ‘best result’, in that it will give the highest number of responses  –  but this approach isn’t cost-free, for a number of reasons:

• Opportunity cost.   In exposing someone to an offer they are not interested in, an opportunity is wasted.  This opportunity cost is in fact always partly borne by the initiator of the action – it’s just not always obvious.     (What if they could have sold the opportunity to someone who could make use of it?)
• Annoyance.    In exposing someone to an offer they are not interested in, you might lose the ability to attract them, at a later time,  to an offer they are interested in.
• Real direct cost.    It’s nice if you can get your users to do your marketing for you, but it’s increasingly the case that you need to get your shovel out and help, too.   When you pay to target your users’ friends, what are you getting for your money?   Do you want all of them, or just some of them?

As connection targeting increasingly becomes a paid-for service, all these types of costs, direct, indirect, and opportunity cost will come under increasing scrutiny.

Let’s look at the study’s results from the other end of the lens for a moment.  If you are like the people in the study (and odds are, you are, as there were millions of users involved), your best friends, from an invitation point of view – i.e. the ones who issue invitations that you are most likely to accept – come from friends who send relatively few invitations,  and send them incrementally, but persistently.     One interpretation of this finding is that these type of invitations are ones that result from friends using their own intelligence and applying it to the developer’s problem.     It may become more important to explicitly encourage this type of accuracy, and value-add, by the type of invitation which is made, and how it is monitored and managed.

What’s the take-away?    Unfortunately, you can’t simply grab the first result you run across, and ride off into the sunset with it whooping and hollering with joy.     You might end up riding in the wrong direction.   After all, the type of product that is being recommended has an influence how people behave when recommending it, as shown by Leskovec, Singh and Kleinberg in their 2006 paper, Patterns of Influence in a Recommendation Network.        And there are other things to think about too.

However, there is one moral that can be easily squeezed out of the results, which is that there is a lot to learn from asking the question.  You don’t ask, you don’t get applies to behavioural insights as well as to lots of other things in life.