Are all Influencers Created Equal?

[#1]

Dr. Michael Wu, Ph.D. is Lithium's Principal Scientist of Analytics, digging into the complex dynamics of social interaction and online communities.

He's a regular blogger on the Lithosphere and previously wrote in the Analytic Science blog.

You can follow him on Twitter at mich8elwu.

---

Last week I presented a webcast with Prof Barak Libai and Sanjay Dholakia on the Holy Grail of WOM marketing. A central theme of the webcast is the importance of influencers, because they are essential in driving the success and effectiveness of WOM in online communities. And in terms of ROI, seeding a WOM program with influencers is 45% more effective than seeding randomly.

I presented one way of identifying influencers using Google's PageRank^TM algorithm. But PageRank is only one of many social network metrics that we may use to score community members; we can do much more. Although PageRank may sound like something novel in the hi-tech world, it is actually a variant of the eigenvector centrality, which has been used in social network analysis (SNA) since 1970s. Let me briefly discuss the subject of social network analysis and what it means to markers.

As alluded to earlier, SNA provides numerous centrality measures that quantify the importance of individuals in a network. There are four centrality measures that are popularly used in SNA and each quantifies importance in a different way.

1. Degree centrality: measures how many connections a user has.
2. Closeness centrality: measures how fast a user can reach the whole network.
3. Eigenvector centrality: measures how reputable a user is.
4. Betweenness centrality: measures how many critical diffusion paths go through the user.

Depending on the desired result and the resource constraint of your marketing campaign, the optimal selection criteria for seeding your WOM program will differ.

If you simply want to spread your message to as many people as possible, you want to seed your program with members that have high degree centrality. On the other hand, if you want to spread your message as fast as possible (for example, if you want to reach 100K people in the shortest amount of time), then you will need to target members with high closeness centrality. If you care about conversion, and you must choose influencer that are reputable. In that case, you should pick members who have high eigenvector centrality. Finally, if your resource is very limited, (say you can only give away one car for someone to test drive) then you need to find the members that has the highest betweenness centrality.

Therefore, influencers are not created equal! But each type has its strength. And it is crucial to understand that influencers who can spread your message to the most people are not necessarily the same one who can do it fastest. Influencers that have the most friends are not necessarily the most reputable. Although in practice these centrality measures are somewhat correlated, understanding the science behind influencer selection and when to use which criteria will maximize the ROI of your WOM programs.

[#2]

Good stuff, Michael! Which of these algorithms is implemented (or planned) in the Lithium Analytics offering? Thanks!

[#3]

Hello MikeTD,

Thanks for the question.

In fact, I've implemented all 4 of the above centrality measures in addition to the PageRank score. We also have several other less well known statistics for each user. And these include:

1. Clustering Coefficient, that measure cliquish a user is. A user that is very cliquish is one where all their friends know each other. A user that is not cliquish will have many friends that do not know each other. So clustering coefficient essentially gives the probability that any 2 of your friends are also friends.

2. Potential Reach, that measure how many user they can reach within 2 degrees of separation. Remember, degree centrality measures how many connections a user has; that is the number of user you can reach in 1 degree. Since in approximately 6 degrees, you can pretty much reach everyone in the world. So number of users you can potentially reach in 2 degrees is what I defined to be a user's potential reach.

I've also implemented 2 statistics that is of interest mathematically for research purpose.

1. Core Number, that measures how central a user when all less central users are removed.

2. Vertex eccentricity, that measures the longest distance between a user and the rest of the network.

These may be useful if they turn out to be correlated to user behavioral characteristics that are relevant to the business. For example, I suspect that core number might be correlated to a user's credibility. If research confirms this, then we can have a user credibility score that is based on the core number. But that is TBD later.

Currently, our analytics offering has an influence score that is based on the PageRank algorithm, degree centrality and potential reach. Other social network metrics are already computed and stored in our data warehouse, and they can be made available any time should we decide that there is a use or demand for it.

[#4]

Very interesting to read about the types of centrality. This is great stuff. I'm definitely going to keep reading your posts.

@socialnerdia

[#5]

Thanks Estaban for you comment.

There are much research to be done along these areas. Social analytics is such an new territory with so many unknown variables and so much to be explore. That is one of the reason that I am so excited about it.

[#6]

Thanks for the response and the link,

My concern is that Lithium is not leveraging as much as possible the differentiation that comes from all these metrics.  It could be that your customers don't fully understand them or implement them, or that they are overwhelmed beyond the basic numbers -- but , as you say, there is so much cool stuff that users could be looking at to make decisions that affect the business.

My comments on the CHI were somewhat vague because I am not certain that you cannot do the things I was calling for, but because I don't think that either users figured it out or that there is a simple way to do that.

Which is why i'd love to see the CHI or another model start working deeper in reputation - not just influence -- which is becoming more critical as the content is created more and more with "unknown authors".  Reputation can bring some balance to the value of the user-generated and community-generated content.

Just two cents, and I am definitely looking forward to read more about it -- and see more implementations centering on reputation, not as much influence... too much to ask?

[#7]

Hello Esteban,

Thanks for the comment. This work on Influencer is currently not related to CHI yet. As we know more about how to evaluate the customer lifetime value (CLV) of influencers, then it can be incorporated into the next version of CHI (which may take a different name).

But the next version of CHI will take into account of the business objective, which we can coarsely segment into 3 type: 1. promote brand and drive sales, 2. support customer, and 3. innovate product. Currently CHI is only a measure of how well the community serve the end-users, but does not take into account of how well does it meet the business objective of the sponsoring company.

I totally agree that reputation will play an important part in the final calculation, because on a community, users usually do not have a pre-established relationship, so reputation is a very important factors that determins a user's influence. That is very difference from a social network (Facebook, LinkedIn, etc) where people usually have pre-established relationships (either friends, relatives, or co-workers or business partners). So estimating influence and reputation in a community is not so trivial. In fact the pagerank algorithm and the eigenvalue centrality mentioned aboved, is an attempt to measure the authority and reputation base on the network structure of who interacts with who.

I hoep this gives you an idea of where all these are heading. Just taking one steps at a time.