Andreas Weigend, Social Data Revolution | MS&E 237, Stanford University, Spring 2011 | Course Wiki

Class_07: Models of Social Influence

Date: April 19, 2011
Audio: weigend_stanford2011.07_2011.04.19.mp3
Initial authors:[David Kim,], [Ruchi Varshney, rvarshne][Eva Petrova-Ibarria, evapi]


Can we measure influence?

  • Let's take our shot at scoring the influence of some people around us!
  • Please complete survey by end of Tuesday, Apr 26!

Key Points

  • Introduce Machine Learning Techniques - Supervised vs. Unsupervised learning
  • Meet Guest Speaker Matthew Russell, author of our textbook "Mining the Social Web"
  • Discuss Models of Influence with Students and Guest Speaker

Designing Social Systems

The design of social systems has become extremely important in helping companies understand what things are important to users. Such systems also need to present useful functionality or value to the user. Some examples include:

  • The "Like" button on Facebook. It provides users with the ability to express their feelings for a particular page or comment, while it also provides Facebook with valuable information about the things its users like and enables it to make better suggestions or target ads based on users' needs.

  • Friend lists on Facebook. It provides users with a handy feature to organize their relationships, send group or event invitations, and filter updates from certain group of friends. On the other hand, it gives Facebook information about who is important to its users. Facebook is able to suggest which friends you would like to add to a particular list based on group information.

Machine Learning Techniques

  • Humans learn from their past experiences. Computers don't have "experiences". A computer learns from data, which represents some past experiences of an application domain.
  • Machine learning refers to the design and development of algorithms that allow computers to automatically evolve behaviors based on raw data. Based on the desired outcome of these algorithms, machine learning techniques can be broadly classified as supervised and unsupervised learning techniques.
  • More details on machine learning techniques are on this wiki. In fact, the Stanford CS program has a whole course taught by Andrew Ng on this very subject

  • A company out of Stanford called ZunaVision is using computer vision to automate a process in which they splice in advertisements for products into the foregrounds and backgrounds of pre-recorded scenes. This could potentially lead to geo-sensitive and personal context-sensitive ad placements in movies and TV if this gains traction.

Supervised Learning

Supervised learning is based on a known training set. The desired target of the learning algorithm is known. More details on this wiki. Examples:
  • Stock market predictions: the algorithm makes a prediction for a stock, and the result is known in a few days.
  • Facebook friend suggestions: the algorithm tries to suggest a friend to you, and the result is known when you click to add that person as a friend.

Unsupervised Learning

Unsupervised learning refers to the class of problems where the goal is to determine how data is organized. We often like to know how information is clustered. However, interpreting the hidden meaning behind clustering patterns is completely up to us. More details on this wiki

Source: Harri Valpola 2000-10-31

Models of Influence

Most machine learning systems can be easily classified as supervised or unsupervised (e.g. there is clearly an outcome variable or not). However, the problem of determining an inherent trait of a person, such as influence or even intelligence, is such a new area in real world problems, that it is impossible to classify it as either.

There is no ground truth around the modeling of influence - there is no proven way to solve this problem. Can the influence of a person be expressed as a number on a scale? In dollars? In metric units? The idea of "influence" seems intuitive but the difficulty arises in concretely specifying metrics that can serve as proxy for the concept of influence.

There is no hard and fast rule around developing a model that predicts the true influence of a person. Yet people are grappling with people who influence them everyday! This is a tough problem to solve.
  • Food for thought: is an IQ test truly representative of a person's intelligence? What is the model used behind designing the GMAT or GRE scoring system?

Why is modeling influence important?

One may wonder as to why the modeling of social influence is important. If you were a marketeer and were given $1,000,000 to spend on a marketing campaign, what would you do? Would you pay a $1,000,000 for a TV ad slot that shows your product once, or would you pay $1 to 1,000,000 seed users on Twitter to spread the word? Would you pay one strong influencer because he/she has a lot of followers? Or would you try to find a number of relatively good seed influencers to potentially get more retweet cascades?

Influence models can help understand retweet cascade patterns and develop optimization functions around ensuring that marketers get more bang for the buck. Brand campaigns are inevitably looking to maximize their ROI. In fact, in today's world the value of people talking about your product is sometimes higher than the money you would get for it.


4 Steps of Modeling

Modeling is an iterative process. Revisit the PHAME methodology from Class02_03.31
  • Problem formulation
  • Data
  • Algorithm
  • Evaluation

Andreas' Big Question: Can influence be equated to money?

  • Most people can be influenced with money. So can influence be expressed as a dollar amount?
  • Is it possible to map a numerical influence score to a certain amount of money? What would such a function look like?
  • Gary Becker, a Nobellaureate in the field of Economics, has written several papers on social economics.
  • Aside: The field of decision analysis developed largely at the Management Science and Engineering Department at Stanford attempts to facilitate decision making through an introspective process where the decision maker attempts to assign a monetary value to the outcomes of various decisions. Perhaps this same framework could be used to at least begin to think about ways of assigning monetary value to a person's "influence."

Issues with measuring influence with a monetary value

  • The amount of influence depends on the context and so the monetary value will change depending on the circumstance.
  • Some things cannot be divided as granularly as money. For example, how much money would you assign the ability to influence someone to "half" do an event such as taking out the trash.
  • Money is not a great metric for people where an increase in wealth does not necessarily result in an increase in personal utility. Or there are those who refuse to be influenced regardless of how much money they are offered (e.g. for religious reasons)

Matthew Russell

About Matthew

  • Matthew Russell, Vice President of Engineering at Digital Reasoning Systems and Principal at Zaffra, is a computer scientist who is passionate about open source, data mining, and web application technologies.
  • Matthew spent many long nights and weekends writing our book "Mining the Social Web" and hopes we will find it useful.
  • Read about how Matthew defines the Social data here.

Matthew talks about Influence

How do we measure influence on Twitter?

Comments/questions during discussion with the class:
  • Do we measure the influence of an individual by the influence of his/her cluster?
  • If you have followers that are geographically further away, does that imply a further reach in influence?
  • It is important to measure the overlap between your followers. Do you have a tight clique of followers or a more far reaching network?
  • How often are your tweets "mentioned"? You haven't initiated the tweets but others are nonetheless mentioning you. Is this influence?
  • How many times was the user favorited (or "bookmarked")? The more bookmarks on this user the higher his influence may be.
  • How quickly do your tweets generate an action?

Viewpoints on measuring influence beyond Twitter

  • What if we measured influence based on the network structure/topology instead of network activity?
  • Would it be wise for marketers to target centralized people, independent of their prior tweet activity?
  • We could hypothesize trust as a measure of influence – special kind of influence that has real world value as a community. But how would we measure trust?

Other insights from Matthew and Andreas

Gathering Website Statistics

Innovative promotions
  • Kindle promotion - get $25 off when you buy a kindle and agree to receiving ads on the device. But the contradiction here is that companies want to reach people that value their time more than they value their $25
  • Some phone companies let you make a free phone call if you listen to a pre-recorded ad before making the call.

Twitter and Monetary Value
  • How much is a follower worth? Today, you can buy followers on Ebay for less than a penny each. Click here and here to read the full TechCrunch stories.
  • A website, called TwitterValue measures how much your Twitter profile is worth. @BarackObama is worth $41,150
  • Another website, called TwittAd lets you auction off ads on your Twitter page. TechCrunch, which has 26,361 followers, is worth $503 per month in ad revenue.

Example: Propagation of the Bin Laden News on Twitter

external image tweet.jpg
Matthew Russel's work tries and correlate influence with the penetration of an authored tweet within the social graph.

Here is a visualization of how the news of Bin Laden being killed exploded on Twitter. Keith Urban (Chief of Staff to former Secretary of Defence and Hoover Institution Fellow Donald Rumsfeld) was the first credible source to break the information so the activity around him is pretty natural.

However, the activity generated by the New-York Times reporter Brian Stelter probably somehow denotes his influence in the field of politics, as he is the only other big "hub" in the diffusion process.

We can then identify a number of smaller "hubs" down the cascade, that we can interpret as secondary "influencers" in this particular context.

More detailed analysis can be found in the original article on Mashable.

Concluding Thoughts

Everyone has an intuitive sense of what influence means. For example, most would agree that the president is one of the most influential people in the world. Indeed, there seem to be widely accepted benchmarks for "influence" including Time magazine's annual "Most Influential People" list as well as Klout's influence score. And yet, it is a concept that eludes specific numerical quantification and goes beyond the traditional realm of machine learning.
Nonetheless, as the amount of data collected of individuals continues to explode in quantity, we may hope to find a metric that reconciles a more intuitive qualitative sense of influence with our desire for a quantifiable index.

Administrative Issues

Results on HW1

  • Check out the best answers posted by Jason at HW1

Dog Food #3

  • Due Thursday (4/21) at noon
  • Check out Quora and answer the questions posted at DF3


  • Due Tuesday (4/26) at noon
  • You will be given a list of 20 people - these people represent the whole Twitter user base for the purposes of this assignment (not just a random sample)
  • Details will be provided on 10 of these people, including their Klout scores, Alexa traffic ranks, US traffic ranks ( i.e. Yelp has higher traffic in the US than in Europe), In-link data
  • Watch out for Twitter API rates!
  • More details available at HW2

Influential People Survey Results!

Thank you everyone who took the time to fill in our Influential People Survey!
Check out USATODAY rankings on the 25 most Influential People in the world here. This is the list of people who changed our world, transformed technology, mapped the human body and affected the way we relate to one another.

Some results from our survey are listed below. Notice that Influence is all subject to personal interpretation. Yet, in the end, everyone ranked Microsoft Chairman Bill Gates as the most Influential person of our survey. This result aligns well with the USATODAY's rank for Bill Gates, which surely was at #1.