Surfing a web of trust:  Reputation and Reciprocity on CouchSurfing.com Debra Lauterbach, Hung Truong, Tanuj Shah, Lada Ad...
<ul><li>Motivation </li></ul><ul><li>Research questions </li></ul><ul><li>Findings </li></ul><ul><li>Results of predictive...
<ul><li>“ Worldwide network for making connections between travelers and the local communities they visit” </li></ul><ul><...
Whose couch would you stay on?
<ul><li>Online->offline transactions require even more trust </li></ul><ul><li>It’s not always easy to decide who is trust...
<ul><li>Bialski & Batorski (2006) examined which factors contribute to higher trust between CouchSurfing friends </li></ul...
<ul><li>Reciprocity: </li></ul><ul><ul><li>Does reciprocity exist, and to what degree? </li></ul></ul><ul><li>Trust: </li>...
<ul><li>600,000+ users </li></ul><ul><ul><li>City </li></ul></ul><ul><ul><li>Country </li></ul></ul><ul><ul><li>Date they ...
<ul><li>1,500,000+ friendship connections </li></ul><ul><ul><li>Friendship degree (scale of 1-7) </li></ul></ul><ul><ul><l...
<ul><li>Most users have few friends (62% have none) </li></ul><ul><li>A small number of users conduct the majority of surf...
<ul><li>Direct reciprocity: 12-18% of stays are directly reciprocated </li></ul>Reciprocity A B
<ul><li>Generalized reciprocity: large strongly connected component (1/3 of active users) </li></ul>Reciprocity
Worldwide reciprocity
<ul><li>Vouching means you believe that friend to be trustworthy </li></ul><ul><li>You can only vouch for others if you ha...
<ul><li>High rate of vouching </li></ul><ul><ul><li>95% of users with > 10 friends have been vouched </li></ul></ul><ul><u...
<ul><li>Tight web of trust….or vouching too freely? </li></ul><ul><li>Mutual trust….or social pressure to reciprocate? </l...
<ul><li>A high number of vouches are between  “CouchSurfing friends” </li></ul>Who are users vouching? Friendship degree: ...
<ul><li>Can we use the available friendship connection information to predict if a connection is vouched? </li></ul><ul><l...
<ul><li>Two-step indirect measure for propagating vouches: </li></ul>Predicting vouches - global measures A B C D ? Indire...
<ul><li>Results from logistic regression for each variable alone: </li></ul><ul><li>Global measures are poor predictors of...
<ul><li>We examined the properties of a worldwide network used to facilitate offline interactions. </li></ul><ul><li>Frien...
<ul><li>Questions? </li></ul><ul><li>Debra Lauterbach [email_address] </li></ul><ul><li>Hung Truong [email_address] </li><...
Upcoming SlideShare
Loading in...5
×

Surfing a Web of Trust: Reputation and Reciprocity on CouchSurfing.com

4,470

Published on

Published in: Technology, Sports, Education
1 Comment
6 Likes
Statistics
Notes
No Downloads
Views
Total Views
4,470
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
0
Comments
1
Likes
6
Embeds 0
No embeds

No notes for slide
  • The study I will tell you about today is largely descriptive, where we focused on finding and interpreting the phenomena occurring in this unique online community, and in particular in the reputation system in use on the site So today in my talk I will cover… ..wrap up with some discussion of how the best practices from CS could be applied to other online communities, &amp; our ideas for future work
  • So, what is CouchSurfing? I’ll give a short introduction as many of you may not have heard of it before - it’s a popular but a niche website. And to give you an idea of how it works, I’ll give you an example. So here we are, most of us visitors in Vancouver. Instead of staying in a hotel or wherever you are staying, you could have used CouchSurfing to find a local person to stay with. You would have done a search for Vancouver, and you would have found that there are over 100 people in Vancouver to choose from.
  • Here are just a few of those CouchSurfing members here in Vancouver with available couches. There is no monetary exchange involved, the whole idea and mission behind the site is to meet new people and share cultural experiences. So, who would you choose, and would you do it?
  • For many people, their answer depends on how much they feel they can trust the other person. You want to know when you’re staying with someone that you will be safe in their house, they will not hurt you - all the basic human survival instincts. Things like the personality and interests of the other person may matter to you too as to how much you may have in common and enjoy meeting each other, but that’s all secondary. Now, it’s already known that trust is a big factor in enabling successful online transactions. Research on eBay, for instance, has shown the importance of their reputation systems for creating trust between the buyer and the seller. So you can only imagine that on sites like CouchSurfing that exist to facilitate offline transactions that trust is even more important. Much research has been done on trust and reputation systems, especially on eBay and other marketplaces, P2P networks, etc. But little research on sites like CS where people meet in person. This is kind of the new frontier of social computing - bringing the experiences into the real world, so studying how this process is faciliated on a real life community was our main motivation.
  • They analyzed a set of private trust ratings, not the public reputation system that we’ll be focusing on - higher trust most correlated with origin and context of relationship, and duration reciprocity was the other salient factor to the site’s success, and the other thing that a reputation system must help enable. Otherwise, not enough hosts to sustain activity
  • In what FORM it exists Examined reputation system from several angles
  • Anonymized
  • This is where it gets interesting…CS contains all these details on friendships, to help users judge others based on their friendships. They are also unique in that their social norms really discourage having many weak friendships like you might see on other sites, so you typically only add friends if you are actually friends in real life.
  • So we did a thorough analysis of the data, beginning with looking at the whole network. And quickly, what we can see is that activity is distributed the way you see on many other sites.
  • Direct: we think this shows that personal connections are being made because of surfing experiences
  • Generalized: means that I may host you, then you may not host me but may host someone else. What you get when this happens is a strongly connected component in the network, where you can hop from couch to couch between users who have surfed or hosted together to every other user in that component. In this case, this component makes up 1/3 of active users who are those who have surfed or hosted at least once
  • We looked at where these experiences were taking place, and you can see that there is a lot of global activity. The size of the circle is the number of experiences within that country, and the lines are experiences between people from different countries. A network as broad, global, and active- must be supported by an underlying network of trust. so, turn attn to trust system
  • The part of the reputation system we studied was the vouching network.
  • How is the vouching system being used? While as I said before only 6.8% of users have been vouches, this is a bit misleading as it is skewed by the low activity on the site Vouches given out (by those who can)
  • From these results, we wondered… Pressure since vouching is public
  • CS friends make up many of the vouched edges
  • Wanted to quantify the factors that contribute in one person vouching for another. We took all the connections that were vouched, along with any connections that had the potential to be vouched but were not , and built a balanced sample of 100,000 of these edges. We then ran a logistic regression model using 10 fold cross validation in order to discern which variables would most accurately predict a vouched edge
  • Surfing a Web of Trust: Reputation and Reciprocity on CouchSurfing.com

    1. 1. Surfing a web of trust: Reputation and Reciprocity on CouchSurfing.com Debra Lauterbach, Hung Truong, Tanuj Shah, Lada Adamic University of Michigan School of Information
    2. 2. <ul><li>Motivation </li></ul><ul><li>Research questions </li></ul><ul><li>Findings </li></ul><ul><li>Results of predictive modeling </li></ul><ul><li>Discussion & future work </li></ul>Outline
    3. 3. <ul><li>“ Worldwide network for making connections between travelers and the local communities they visit” </li></ul><ul><li>1,300,000+ members </li></ul><ul><li>231 countries </li></ul>About CouchSurfing
    4. 4. Whose couch would you stay on?
    5. 5. <ul><li>Online->offline transactions require even more trust </li></ul><ul><li>It’s not always easy to decide who is trustworthy </li></ul><ul><ul><li>This is why having a reputation system is key </li></ul></ul>The importance of trust
    6. 6. <ul><li>Bialski & Batorski (2006) examined which factors contribute to higher trust between CouchSurfing friends </li></ul><ul><li>Molz (2007) examined the sociological meaning of reciprocity in the context of hospitality exchanges </li></ul>Related work
    7. 7. <ul><li>Reciprocity: </li></ul><ul><ul><li>Does reciprocity exist, and to what degree? </li></ul></ul><ul><li>Trust: </li></ul><ul><ul><li>Who is doing the vouching? </li></ul></ul><ul><ul><li>Who is being vouched for? </li></ul></ul><ul><ul><li>Can we predict which connections are vouched? </li></ul></ul>Our research questions
    8. 8. <ul><li>600,000+ users </li></ul><ul><ul><li>City </li></ul></ul><ul><ul><li>Country </li></ul></ul><ul><ul><li>Date they joined CouchSurfing </li></ul></ul><ul><ul><li>Number of profile views </li></ul></ul>The data
    9. 9. <ul><li>1,500,000+ friendship connections </li></ul><ul><ul><li>Friendship degree (scale of 1-7) </li></ul></ul><ul><ul><li>Whether they met in person (Y/N) </li></ul></ul><ul><ul><li>How they met (through CS, or offline) </li></ul></ul><ul><ul><li># days traveled together </li></ul></ul><ul><ul><li># days hosted </li></ul></ul><ul><ul><li># days surfed </li></ul></ul><ul><ul><li>Whether they vouched (Y/N) </li></ul></ul><ul><ul><li>Rating of their overall experience (-1 to 1 scale) </li></ul></ul><ul><ul><li>Date friendship connection was made </li></ul></ul>The data
    10. 10. <ul><li>Most users have few friends (62% have none) </li></ul><ul><li>A small number of users conduct the majority of surfing/hosting experiences </li></ul>Patterns of activity
    11. 11. <ul><li>Direct reciprocity: 12-18% of stays are directly reciprocated </li></ul>Reciprocity A B
    12. 12. <ul><li>Generalized reciprocity: large strongly connected component (1/3 of active users) </li></ul>Reciprocity
    13. 13. Worldwide reciprocity
    14. 14. <ul><li>Vouching means you believe that friend to be trustworthy </li></ul><ul><li>You can only vouch for others if you have at least 3 vouches yourself </li></ul><ul><li>Vouching forms a small “web of trust” in the network </li></ul><ul><ul><li>6.8% of users have been vouched at least once </li></ul></ul><ul><ul><li>1.8% can vouch for others </li></ul></ul>Reputation System - Vouching “ Respecting the significance of vouching is essential to the integrity of the network... It is very important that you ONLY vouch for people that you have met in person and know well enough to believe that he or she is trustworthy.”
    15. 15. <ul><li>High rate of vouching </li></ul><ul><ul><li>95% of users with > 10 friends have been vouched </li></ul></ul><ul><ul><li>15.07 vouches given out on average </li></ul></ul><ul><ul><li>25% of friendships that can be vouched are </li></ul></ul><ul><li>High rate of reciprocity </li></ul><ul><ul><li>74.6% of vouches are reciprocated </li></ul></ul>Is the vouching system being used as intended?
    16. 16. <ul><li>Tight web of trust….or vouching too freely? </li></ul><ul><li>Mutual trust….or social pressure to reciprocate? </li></ul>Why do users vouch others?
    17. 17. <ul><li>A high number of vouches are between “CouchSurfing friends” </li></ul>Who are users vouching? Friendship degree: 1= Haven’t met yet 2= Acquaintance 3= CouchSurfing friend 4= Friend 5= Good friend 6= Close friend 7= Best friend
    18. 18. <ul><li>Can we use the available friendship connection information to predict if a connection is vouched? </li></ul><ul><li>Logistic regression model (10-fold cross-validation) </li></ul><ul><li>71% accuracy in predicting whether a random edge is vouched </li></ul><ul><li>Most predictive attributes were friendship degree, rating of experience, how they met </li></ul>Predicting vouches
    19. 19. <ul><li>Two-step indirect measure for propagating vouches: </li></ul>Predicting vouches - global measures A B C D ? Indirect vouch score for A->D: = 1/n(B) + 1/n(C)
    20. 20. <ul><li>Results from logistic regression for each variable alone: </li></ul><ul><li>Global measures are poor predictors of whether an edge is vouched </li></ul>Predicting vouches - global measures Predictive accuracy: Variable 50.6% PageRank 54.2% 2-step vouch propagation 55.8% Jaccard coefficient 67.7% Friendship degree
    21. 21. <ul><li>We examined the properties of a worldwide network used to facilitate offline interactions. </li></ul><ul><li>Friendship degree information is beneficial </li></ul><ul><li>Global measures may be useful in assigning overall reputation scores, but not for predicting if a specific person will vouch for another or not </li></ul><ul><li>Further work is needed to determine if vouches are given too freely </li></ul>Summary
    22. 22. <ul><li>Questions? </li></ul><ul><li>Debra Lauterbach [email_address] </li></ul><ul><li>Hung Truong [email_address] </li></ul><ul><li>Tanuj Shah [email_address] </li></ul><ul><li>Lada Adamic [email_address] </li></ul>Thank you!

    ×