Adaptive Influence Model: Fixing the Influence Irony

[#1]

Last time we took a quick peek at the history of SEO, and we saw that influence engine optimization (IEO) is an inevitable consequence of scoring people’s influence. What’s worse is that IEO leads to the influence irony, where it actually changes people’s behavior in a way that drives them further away from being truly influential (if you missed this crucial point from my last post, you should read The Influence Irony – Influence Engine Optimization).

This sounds disappointing, but today, we are going to fix it!

If you look back at the Webster’s definition of influence that I quoted earlier (see What is Influence, Really? – No Carrots, No Sticks, No Annoyance, No Tricks), it says: “The power or capacity of causing an effect in indirect or intangible ways : sway.” So, whatever the effects are (i.e. the changes in thought or behavior), it must be caused through “indirect or intangible ways.”

What does this mean? It is actually a subtle but important point that many discussions about influence have overlooked. It means that influence shouldn’t be something that people can affect directly and easily. This is the key to fixing the influence irony. Because people would otherwise be able to game the system and raise their own influence score without doing anything influential.

Why is the “Influence Irony” so Challenging?

To mitigate the gaming of influence, vendors must ensure that people can’t directly affect their influence. However, influence should be something that people can affect through their actions. Otherwise people’s influence will never change. So influence is something that people can affect, but not directly or easily.

This subtle point of influence is already very hard for people to grasp, but if you think deeply about it, it sort of makes sense. If you can directly change someone’s thoughts or behaviors, then you are no longer influencing them. You are controlling them, and that is not influence. Are you following me?

Now, the real challenge comes when you try to make this rigorous enough so you can write an algorithm to compute someone’s influence score. Part of the difficulty is because this statement is quite vague: “influence is something that people can affect, but not directly or easily.” How direct is considered indirectly, and how easy is not easily? What are the criteria to determine if something is too easy or too direct?

Even though we can understand this in words, it’s actually very hard to put it into equations and programming codes, and the precise criteria are very hard to determine. But luckily there is a solution! The solution is to leave it up to the crowd and let the population determine the criteria for you. This is the essence of an adaptive model.

What is an Adaptive Influence Model?

First let me clarify the difference between an adaptive model vs. an adaptive score:

• An adaptive influence score means that the score is adaptive to the behavior of the individual. Most influence algorithms already do this by incorporating a timing factor in their influence scoring algorithm. This is easy, because when an individual’s behaviors changes, his social media activity data will also change to reflect his behaviors. Since each person’s influence score is computed from his social media activity data, the score will obviously change, because the input data to the model has changed
• An adaptive influence model means the algorithm that calculates the score will change and adapt to the behaviors of the population. Most influence vendors do not do this, because it is much harder. However, it is this evolution of the model that is going to fix the influence irony. Let’s take a look at what this means in greater detail

How to Adapt your Influence Model to the Population?

The best way to understand how adaptive models work is to build one. So I will try to walk through the essence with you using an example that we’ve discussed earlier concerning follower count. When we applied the definition of influence, we came to the conclusion that followers shouldn’t affect someone’s influence score, because too many people will just follow back blindly as a quid pro quo to someone following them.

Rather than eliminating the follower metric from the model completely, adaptive models will still make use of this data, but the model will ensure that follower metric doesn’t affect the final influence score (e.g. by assigning a weight of zero to the follower metric). You may wonder why bother using the follower metric if we want to ensure that it has no effect? This is the essence of an adaptive model, because now we can use behavior feedback data from the population to adjust this weight.

If everyone (i.e. 100% of the population) is following back blindly, then follower count is merely a reflection of the quid pro quo behavior, which is not influence. In this case the weight should be set to 0. On the other hand, if one day people become very selective of whom they follow and only follow people who produces good content that affects their everyday decision, then follower count should contribute to someone’s influence score. In this case, the weight should be set close to 1. In an adaptive model, we dynamically adjust the model (e.g. the weight for follower count, or the specific formula and algorithm that we use the compute influence scores) based on the collective behavior of the population.

If you’ve tried to develop an influence scoring algorithm, you have probably analyzed a lot of data and convinced yourself that reciprocity metrics such as retweets or likes are good metrics for estimating someone’s influence score. What happens if people’s behavior change tomorrow?

If everyone starts to retweet everything they received and liking everything they see because they get points and badges for doing that (via gamification), would retweets and likes still be good metrics to use in your influence model? They may still be important, but they should probably be weighted down significantly. Therefore, 2 extra steps are needed to make an ordinary influence model adaptive:

1. We need to monitor what fraction of the population are tweeting/liking because the content is good, and what fraction of the population are retweeting/liking due to gamification
2. Then we need to adjust the weight of the retweets/likes data according to this fraction

Why are Adaptive Models so Hard to Build?

Adaptive models are very hard to construct, because every metric that is used by the algorithm can be gamed. Furthermore, there are probably many ways to game a particular metric. Thus, for every metric the algorithm uses, there are probably 10 or 20 auxiliary metrics the system needs to monitor in order to determine whether people are indeed gaming a metric.

Regardless of the challenges, this is exactly what search engines are doing to prevent SEO experts from gaming their relevance ranking algorithm. As soon as a significant amount of people have figured out a scheme to game a metric used in search ranking, the algorithm will quickly adjust itself to reduce the impact of the gamed metric on the final ranking. Since search engines make use of hundreds of metrics for their relevance ranking, they need to track tens of thousands of auxiliary metrics in order to adapt their algorithm to the population. To make things worse, popular search engines (e.g. Google, Bing, etc.) are also hot targets for this type of gaming (i.e. SEO) behavior, so their ranking algorithms are literally changing every day. After almost 15 years, Google is still adding metrics to their system to improve its ability to adapt to the searching and browsing behavior of its users.

Most influence scoring algorithms uses fewer metrics (i.e. less than 50 metrics) than search engines, so the number of auxiliary metrics these algorithms need to monitor is also fewer. However, as we discussed previously, influence scoring algorithms are not only more susceptible to gaming, they are also easier to game. They need to adapt faster in order to counteract the negative effect of the influence irony.

Conclusion

So we have good news and bad news. The bad news is that the influence irony is an inevitable consequence of scoring people’s influence. But the good news is that there is a way to fix this problem through adaptive models. Although building an adaptive model is challenging, it can be done. The search engine industry has demonstrated its feasibility at the planetary scale. The fact that search engines are still working pretty well is an indication that their relevance ranking algorithms is able to adapt to the gaming behavior of the SEO industry. Otherwise, search engines would return the most SEO optimized page with respect to a query rather than the most authority and relevant pages. And search quality would significantly degrade as SEO continue to game their ranking algorithm.

Currently, no influence vendor uses adaptive algorithm to score influence. But influence is a very nascent industry, and there is a lot for them to learn from the more mature search engine industry. The outlook of the influence industry may be very bright (i.e. where influence scores truly reflect people’s capacity to influence) or very gloomy (i.e. where influence scores only reflects how effectively people game the scoring algorithm). Which future becomes reality is going to depend on how quickly they can implement an adaptive model of influence.

Alright, next time I am going to roll the dice and see what topics I end up writing. Or you can tell me what topics you like to hear. In the meantime, I welcome any comments, discussion, kudos, critiques, and/or challenges. See you next time.

---

Michael Wu, Ph.D. is Lithium's Chief Scientist. His research includes: deriving insights from big data, understanding the behavioral economics of gamification, engaging + finding true social media influencers, developing predictive + actionable social analytics algorithms, social CRM, and using cyber anthropology + social network analysis to unravel the collective dynamics of communities + social networks.

Michael was voted a 2010 Influential Leader by CRM Magazine for his work on predictive social analytics + its application to Social CRM. He's a blogger on Lithosphere, and you can follow him @mich8elwu or Google+.

[#2]

A topic myself and and many others may be interested in. Measurment of Collaboration internally in an organization that is using a collaborative platform ie. IBM Connections. But here is where it gets tricky...you only have the ability to count things like below.

[Blog_Entries]
[Blog_Comments]
[Board_Entries]
[Board_Comments]
[Forum_Posts]
[Wiki_Pages_Owned]
[Wiki_Pages_Updated]
[Wiki_Comments]
[Files]

getting this specific data is easy. Weighting the data is very tricky.

[#3]

Hello BB2,

Thanks for the comment and glad to hear that this is a topic that interests you.

Measurement of collaboration and measurement of influence is completely different thoiugh. I have actually did quite a bit of research in both.

First, no one says you have to take the data as it is. You can transform the data and create new data derivatives (or features, in machine learning and statistics jargon) that are more suitable for your problem. For example, simple data derivatives using the list of variables you provided could be computed by taking a ratio of 2 variables:

     [Blog_Comments] / [Blog_Entry]

This new data derivative can be a better indicator of collaboration than the 9 variables you provided. But it may not be the best. You can also create more complex features like:

     [Number_of_Unique_Participants] / [Board_Entry]

If you are an experienced machine learning or statistician, you would create even more complex features based on networks dynamics, time series, etc. And in fact, from my research on quantifying collaboration, the most effective features are communication network based features. And I've presented some of these findings in several venues. One of these presentation is at E2, and it is captured in video here:

http://www.informationweek.com/thebrainyard/video/1697502738001

Please note that this is really a 1 hour talk that's pack into 30 minutes, so there are a lot of details that I've left out in this talk. If you are interested, please feel free to ask questions about this talk here and I'm happy to expand the discussion here.

Hope to see you next time.

Thanks.

[#4]

Gamification influences people to participate more then people gamify their behaviour to be (seen as) more influancial. Is that right Michael ?

[#5]

Michael, I want to express a practical solution to your series of articles around measuring influence. I have read along in recent months and I have been waiting for this point of view to come up in conversation -- as it has not I would like to introduce the thread.

It seems to me that an immediate and practical solution to measuring the change in thought or behavior of the "influencees" is to track Performance of Influencers who participate in marketing programs.  How well do they drive action or key results on behalf of the program?

In terms of the practical application of these theories, I submit that we can account for the majority of metrics required to deliver the real influence measurement by tracking social outcomes from brand marketing programs --  by default these will incorporate the capacity for influence of the target audience ("influencees").

In other words: If a brand has a community of influencers/advocates that are scored with an algorithm that includes Influencer Performance + Influencee Outcome, haven't we, for practical purposes, solved the problem at hand?

[#6]

Hello Arnaud,

Thank you for the continuing the conversation on this topic from the last article.

I’m not sure if I completely understand your statement. Not sure if it’s the English language that got in the way, but I would say it in a slightly different way and in greater detail.

Gamification can influence people to participate more. However, it depends on what data you are feeding back to the users (Feedback of measurements on human behavior is one of the simplest way to gamify human behavior). It also depends on if more participation is what you want, because participation ≠ influence.

If influence score can be increase by increasing participation (which is pretty easy) regardless of the quality of participation (which much harder), then people will change their behavior to participate more. In that case, basically the influence score is gamifying participation rather than influence. So people will participate more, but their true influence will actually decrease.

I hope this clarify the distinction between several related but subtly different concepts. That is gamification, feedback, and measuring influence.

Alright, thanks again.
See you next time.

[#7]

Hello JohnEaton2011

First of all, thx for taking the time to comment and bring up an interesting point. I apologize for the late response. I am usually much faster. But being totally swamped and getting over the flu can really slow you down.

Your method of measuring influenceis sound theoretically. The precise definition of an influencer’s influence depends on the effect it has on the target influencee’s change in behavior or thought (see Why Brands STILL don't Understand Digital Influence?).

But the challenge is in the implementation. How do you measure the influencer’s performance? The attribution model here is very challenging. I talked about this a bit in my WIRE UK article.  

Quoting a snippet from that article. To truly measure influence, we need to do 4 thing:

1. Track when a desired change (e.g. a purchase transaction) has occurred (this is the easy part)
2. Determine who influenced this purchase (harder, but not impossible)
3. Prove that this purchase would not have occurred otherwise (this is really hard). If this purchase would've occurred anyway, then the influencer didn't actually cause any change. So by definition, this transaction doesn't represent a real influence.
4. Count all the transactions that represent real influence grouped by the influencers (this step is trivial once you have the previous three).

So yes, you can simply measure the total financial outcome of the marketing program and attribute that to the effect of influencer. But this would be an over simplification, since it’s hard to prove that this outcome could not be realized without the influencer. But only then could we truly attribute the financial outcome to the influencer.

I’ve work with several academics (from MIT Sloan, Stern NYU, and Wharton Upenn) to tackle this problem before. And we have developed a method for measuring the financial outcome of an influencer seeding program. The method is call total impact upon removal. It measures the monetary value that a program would lost if you remove a influencer from the program. That would be the monetary value contributed by the influencer. But this is hard to implement in practice too and only work on agent based model, which we use to compute the net gain in customer lifetime value contributed by the influencers.

Not sure if this address your question. It is a tough one, and we work on it before. It is theoretically sound, but just very difficult to implement in practice.

Thanks for your discussion and interest.

I hope to see you again soon.

[#8]

Thank you for your response - I am working hard to apply some of this theory on a couple of projects I have in motion, so i may be guilty of over-simplifying or glossing over details here and there.

In the meantime, you got me thinking over the last few days and my comment expanded into a post.  wp.me/p25GuL-7u   Thanks for the inspiration!

I'll let you know if I manage to turn any of this into action and results in the next month or two.

[#9]

Hello JohnEaton

Thx for the convo. And I'm happy to hear that my work can still be inspirational...   ;-)

I'm glad to hear that you are working on applying some of the stuff we talked about here. Like I said, discussion is always open on any of my blog articles, and I will always respond. So in the process of applying some of the theory, feel free to pop back in any time if you feel the need to discuss this with someone. I'm sure this will be useful for everyone, and we all learn when we discuss.

Good luck and see you next time.