Date post: | 13-Jul-2015 |
Category: |
Technology |
Upload: | amit-sharma |
View: | 133 times |
Download: | 0 times |
The role of social connections in shaping our preferencesUnderstanding sharing and consumption online
Amit SharmaPh.D. Candidate
Dept. of Computer ScienceCornell University
www.cs.cornell.edu/~asharma@amt_shrma
Collaborators
● Dan Cosley, Advisor, Cornell University● Baoshi Yan, LinkedIn● Gueorgi Kossinets, Google● Jake Hofman and Duncan Watts, Microsoft Research● Students
■ Masters: Meethu Malu, Mevlana Gemici, Michael Triche
■ Undergraduate: Yulan Miao
Finding meaning in social data
People express their connection with items in myriad ways on the web
Examples● Hashtag on Twitter● Like on Facebook● Rate on Goodreads● +1 on Google● Favorite on Etsy
How do these activities connect to people’s decisions on items, products, opinions?
Connection between retweeting and influence, liking and buying, sharing and consuming
How items diffuse through social networks: The “holy grail”
Past work has studied:● intrinsic attributes of
items● the influence of certain
individuals
Selection Bias: Studies on online data show that most shares propagate only one level; a tiny fraction get to more than 1 level
● Can our friends’ activities be used to predict ours?
■ You like this because Jeetu liked it.● Can information about our friends’ activities
help us make decisions on items, form our opinions?
■ Amit Sharma and 10 of your friends like this.● Would our friends suggest items that we
would like?■ Jeff Bezos shared this to you.
Three questions for research
Three questions for research
Can our friends’ activities be used to predict ours?
Can information about our friends’ activities help us make decisions on items, form our opinions?
Would our friends suggest items that we would like?
How to design network-aware recommendation models?
How to present social information in system interfaces? dummy
How to include manual shares in recommender systems?
Ego NetworkSubgraph containing a person and her immediate social connections.
FriendAny first-degree connection of a person as defined by a particular social network.
Preference(Partial) ordering over items that helps a person choose items to consume.
Definitions
Part I: Predicting users’ activities based on friends’ activitiesA study using data from Facebook and Twitter
ICWSM 2013
Datasets from Facebook and Twitter
Preferences: Movie and music Likes on Facebook, hashtag usage on Twitter
Data collected from people who gave permission to Facebook apps [Sharma and Cosley 2013b, McAuley and Leskovec 2012]
What would be good measures of preference locality?
● User similarity-based: how similar are people’s activities on items in the ego network versus the full network○ Similarity between users○ Density of the user-item matrix
● Item coverage-based: how widely spread are items in the network○ Number of ego networks an item is a part of○ Comparison with random graphs
User A : [Titanic, Braveheart]User B: [Braveheart, Star Wars]User C: [Star Wars, Star Trek]
Similarity: Jaccard similarity
Sim(A,B) = ⅓ Sim(A,C) = 0 Sim(B,C) = ⅓
Measures of locality: Similarity
Measures of locality: Sparsity
Measured by the density of the user-item matrix
Density = 6/12 = 0.5
Titanic Braveheart Star Wars Star Trek
User A 1 1 - -
User B - 1 1 -
User C - - 1 1
Evidence of locality for all three domains.
Hashtags show higher locality than artists or movies on Facebook.
Measures of locality: Item coverage
Uncovered Ego: Percentage of ego networks that do not contain a given item.Random Item/Ego: Compare uncovered ego of given network with a network constructed by randomizing the item likes between users.Random Friend/Ego: Compare uncovered ego of given network with a network constructed by randomizing a user’s friends.
Hashtags have highest locality.The metrics are divided between artists and movies on Facebook.
Similar locality results for item coverage-based metrics
So far, we have seen aggregate metrics.
How does locality perform on predicting each user’s preference?Consider a 70-30 split between train and test.Two sets of data:● One using only friends (Friends / Local)● One using whole network except friends (Non-Friends /
Global)Algorithms: k-nn, matrix factorizationEvaluated on NDCG metric, widely used in IR and recommender systems.
k-nn recommender based on friends outperforms or is comparable to those using non-friends
Number of friends ~100-500
Number of non-friends ~50k
NDCG for 50-nn using friends, non-friends and the full network. Recommendations from friends are are still comparable to those from the full network.
Typical use case: Using friends + non-friends
Useful for recommender systems are exposed only egocentric slices of the network (e.g. through third party APIs of Facebook and Twitter)
Part II: How social processes work to influence our preferencesA specific example: Social explanations on the web
WWW 2013
How explanation strategies serve as proxies for social processes
Overall Popularity (OVP)
Social Process: Proof
Count of Friends(CFR)
How explanation strategies serve as proxies for social processes
Social Process: Conformity
Social Process: Influence
Random Friend(RFR)orGood Friend(GFR)
How explanation strategies serve as proxies for social processes
Good Friend & Count(GCFR)
Social Process: Conformity and Influence
How explanation strategies serve as proxies for social processes
A user study (N = 237)
● Within-subjects design.● Musical artists recommendation. Chose
artists which users were not aware of.● Participants were exposed to
recommendation accompanied by different explanation strategies.
● Each participant rated a maximum of 30 recommendations.
● At the end, participants also answered a questionnaire.
Example Interface
Pink Floyd
+Social
Explanation
How likely are you to check out this artist?
Likelihood Rating (0-10 Likert)
=
PHASE I
More insights into people's rating decisions
Showing the right friend matters
Popularity matters only if people identify with the crowd
People are differently susceptible to explanation
“I found it most powerful when I could see what friend likes the artist. I know what kind of music my friends listen to and that helps me know if I would like the artist or not."
“If it was a friend thatI did not think I would have similarly music taste too, thenI immediately ruled the artist out...”
"The recommendations that were most convincing to me were the ones thatdisplayed that a decent number of my friends listened to orliked the artist. I often like to hear my friends’ feedback oncertain artists..."
"Me and my friends’ music tastes rarely match up, so I’ve learned to not care about what music my friends like."
More insights into people's rating decisions
Showing the right friend matters
Popularity matters only if people identify with the crowd
People are differently susceptible to explanation
More insights into people's rating decisions
Showing the right friend matters
Popularity matters only if people identify with the crowd
People are differently susceptible to explanation
More insights into people's rating decisions
Showing the right friend matters
Popularity matters only if people identify with the crowd
People are differently susceptible to explanation
Social explanation is a secondary effect
"The albums with the most interesting picture, or interesting name, with a lot of likes. If the name struck me, such as ‘Formidable Joy’, I found myself wondering more.If a lot of my friends liked it, it must be good!"
Based on a combination of these two decision processes, a user evaluates a recommendation.
Pink Floyd User's receptiveness to an explanation.[Effect of Explanation]
User's discernment in music.[Base Decision Process]
Amit Sharma likes Pink Floyd.
Modeling the effect of explanations
Base Decision Process f(x) = A e-Ax
A generative process of influence for explanations
A: Discernment
Base Decision Process f(x) = A e-Ax
A generative process of influence for explanations
A: Discernment
Effect of Explanations mu : Receptivity
sigma: Variability
Base Decision Process f(x) = A e-Ax
A generative process of influence for explanations
A: Discernment
Effect of Explanations
Mixture Model h(x) = a f(x) + (1-a) g(x) a : Rigidness
mu : Receptivity
sigma: Variability
Good Friend strategies show lowest rigidness.
Why is this a likely model? All models show same discernment ~0.4
● The effect of social explanation varies with different strategies and different people. Can be used for personalized explanations.
● Explicitly named friends (influence) more impactful than count of friends (conformity).
● Still, aggregate effects can be modelled. A generative model gives us a window into people’s decision process.
Findings
What is the role of people’s preferences in sharing?
Past research shows that when broadcasting, people tend to share only highly liked items [Sharma and Cosley ‘11, Naaman et al. ‘10]A lot of sharing still directed at specific people. How do people choose items to share directly with a recipient?● Altruism suggests that people will share what they
expect the recipient to like● Individuation suggests that people will share what they
like themselves
Where does the balance lie, and how can we model it?
A paired study (N=87 pairs)
● Facebook users invite a friend to take part in the study
● See identical recommendations sourced from each users’ movie Likes
● Recommended can be rated and/or shared with the partner
● To control for social influence, users do not know which items were shared to them
Three groups of participants
● Both_shown: Pairs who saw a mix of recommendations personalized on both partners’ Likes○ Own_algo: Personalized for partner A○ Other_algo: Personalized for partner B
● Own_shown: People who saw only recommendations personalized for them
● Other_shown: People who saw only recommendations personalized for their partner
Partners of Own_shown are in Other_shown and vice-versa.
Individuation seems to dominate, but still participants claimed they were personalizing for the recipient
“Usually when I suggest, it depends on the item, notthe target individual, because I want to share what I enjoyed.” [P8]
“I make suggestions to people if I think they might gainenjoyment. Obviously it really depends on their personalityand their likes/dislikes.” [P22]
Preference-Salience model of sharing
People do not really try to balance individuation and altruism when they share items. Rather, they share based on their preference for items and what is salient to them at the moment. Recipient help decide whether to share an item or not.
Alternative hypotheses:High Quality Sharers: Shared items not significantly higher rated on IMDB than non-shared items.Misguided Altruists: Shared items have consistently higher rating by the senders.
People’s decisions on items depend on both preferences and social factors.
Requires mixed methods approach (Data mining + online experiments).
Models of people’s decision processes can predict what items are more likely to be adopted or shared.
thank you
Amit SharmaDept. of Computer Science
Cornell Universityhttp://www.cs.cornell.edu/~asharma/
@amt_shrma