230 likes | 355 Views
“Make New Friends ,but Keep the Old”-Recommending People on Social Networking Sites. Jilin Chen ,Werner Geyer ,Casey Dugan ,Michael Muller , Ido Guy CHI 2009. Outline. Introduction Data Set Algorithm Experiment Personalized survey Controlled field study Discussion & Conclusion.
E N D
“Make New Friends ,but Keep the Old”-Recommending People on Social Networking Sites Jilin Chen ,Werner Geyer ,Casey Dugan ,Michael Muller ,Ido Guy CHI 2009
Outline • Introduction • Data Set • Algorithm • Experiment • Personalized survey • Controlled field study • Discussion & Conclusion
Introduction • Users in online social network site has two type of friends • Already known offline • New friends they discover on the site • There are many personalized-recommended algorithms , but the effective of those approach is not available • It is different from traditional recommendations of books, movie, restaurants, etc.
Introduction • Goal • Effectiveness of different algorithms • The characteristics of recommending known versus unknown people • If the recommender system effectively increase the number of friends a user has • Overall impact of a recommender system on the site
Data Set • online social network site : Beehive within IBM • Start time: July 2008 • Network situation in experiment: 38000 users, average of 8.2 friends per user. • Friend type: Non-reciprocal friendship
Algorithms • People recommendation algorithms • Content matching • Explanation: common keywords • Content-plus-link(CplusL) • Explanation: common keywords & directional links • Friend-of-Friend(FoF) • Explanation: common friend list • SONAR • Explanation: all relation in database of IBM
Algorithm-Content matching • Motivation : If we both post content on similar topics, we might be interested in getting to know each other. • Formulation(similarity of two users) : • Relationship explanation : show up 10 highest scores words.
Algorithm-Content plus link • Motivation: By disclosing a network path to a weak tie or unknown person, recipient may be more likely to accept it. • Link rule(3 and 4 path): • Similarity scores: if valid link exits ,boost 50% • Relationship explanation : show up 10 highest scores plus valid links if it exits.
Algorithm-Friend of friend • Motivation : If many of my friends consider Alice a friend, perhaps Alice could be my friend too. • Formulation: • Score : Number of Mutual friends. • Relationship explanation : show up all mutual friends.
Algorithm-SONAR • SONAR system : Aggregates social relationship information from public data sources within IBM • Organization chart • Publication database • Patent database • Friending system • People tagging system • Project wiki • Blogging system
Experiment :Personalized survey • Methodology: • 500 active users • Every user was exposed to all four algorithms • Top 10 recommendations of four algorithms
Experiment :Personalized survey • For each recommendation , we show a photo, the job title and the work location ,as well as the explanation generated by a algorithm. • User answer following Question for the test.
Experiment :Personalized survey • User also answer more general questions like their interest in meeting people on the site. • 415 logged in and 230 valid survey form. • Results-Understand user’s need • 95% of the user considered people recommendations to be useful and would like to see them as a feature on the site. • 61.6% said they are interested in meeting new people , 31% said maybe and 7.4% say no.
Experiment :Personalized survey • What may make people to connect to unknown person : 75.2% chose common friends , 74.4% said common content, 39.2% indicated geographical location of the person, 27% said the division within IBM, and 14.5% chose “other”.
Experiment :Controlled field study • Methodology: • 3000 users • Divide into 5 groups, each with 600 users.4 experiment with one algorithm, 1 control group that did not get any recommendations. • In experiment group ,show one recommendation a time, starting from the highest ranked ones. • In control group, we advertised various friending features and actions.
Experiment :Controlled field study • Valid users: 122 from content matching group, 131 from the content-plus-link group , 157 from the friend-of-friend group, and 210 from the SONAR group. • Test situation:
Experiment :Controlled field study • In contrast to survey, the introduction response is less than 1% • “what is this” let the users feel bothered and ignore the feature • Impact of people recommendations • In experiment group viewed 13.7% more page compared to previous time • In control group viewed 24.4% less page compared to previous time
Discussion and conclusion • The result can show the four algorithm are effective in making people recommendation and increase the number of friends. • Relationship-based algorithms are better at finding known one ,whereas content similarity algorithms are better at new friends • To combine the strengths of both type of algorithms, we can initially use R-B algo ,complement them with C-S algo latter.