Multi-Layered Friendship Modeling for Location-based Mobile Social Networks Nan Li and Guanling Chen Department of Computer Science, University of Massachusetts Lowell July 14, 2009 Toronto, Canada MobiQuitous 2009
Online Social Network Success Popular (half billion ww users) Sticky (26m per day)
OSN Goes Mobile Already top Web destinations on smartphones Unique feature – location GPS-enabled phones Sharing current location Attaching location to user-generated content Outlook LSN >$3.3B revenue by 2013 (ABI) Dodgeball, Loopt, Brightkite, Whrrl Google Latitude, Foursquare
Brightkite Startup founded 2005, Denver CO Angel funding $1M, 03/2008 Private beta, 04/2008 Opened to public, 10/2008 User activity Check in, status update, photo upload All attached with current location Updates through SMS, Email, Web, iPhone… Social graph with mutual connection See your friends’ or local activity streams
Usage Snapshot
Contributions Data collection from Brightkite 19k users; 1.5m updates Quantitative correlation model for friendship User tags, social graph, location/activity Evaluation using 10m training data and 45d test data Outperformed than Naïve Bayes classifier or  J48 decision tree algorithms
Data Collection Brightkite Web APIs 12/9/08-1/9/09:  18,951  active users Back traced to 3/21/08:  1,505,874  updates Profile: age, gender,  tags , friends list Social graph:  41,014  nodes and  46,172  links Testing data: next 45 days had  5,098  new links added
 
Tag Cloud
Basic Approach Coming up metrics that Differentiate friends and non-friends Tags, social graph, location, activities Combination of the metrics Training and testing with traces
Using Metrics
Metric Combination
Social Graph
Social Graph Metric
Tag Graph 1000 most popular tags as the nodes Complete graph Link weight reflects likelihood of two tags shared by friends
Tag Graph Metric
Location Graph
Location Graph Metric
Rank Value Result
Modeling Accuracy Take another 100,000 non-friend pairs Not in training data Plus the newly added 5,098 friend pairs Sort the prediction values
ROC Curve
Top Recommendations
Information Gain Worldwide buzz: Planetary-scale views on an instant-messaging network. J. Leskovec and E. Horvitz, June 2007.
Discussions Model stability as Brightkite grows Does not require frequent re-calculation On-demand recommendation Heuristics to speed up metric calculation Possible improvement Different metrics, or combination methods “ Private” updates Conjectured to be few, but no proof
Related Work Industrial solutions: Facebook, Twitter Technical details unknown OSN structural analysis Aggregated behavior not suitable for individual recommendations Collective filtering User-item vs. user-user
Conclusion Correlated attribute combination has good friendship recommendation power Interests, social graph, location Location metric is important Gender and age not so much Future work System implementation Real-user action-based evaluation
Acknowledgement Anonymous reviewers Shepherd- Sharad Agarwal Best Paper Award committee

Multi-Layer Friendship Modeling for Location-Based Mobile Social Networks

  • 1.
    Multi-Layered Friendship Modelingfor Location-based Mobile Social Networks Nan Li and Guanling Chen Department of Computer Science, University of Massachusetts Lowell July 14, 2009 Toronto, Canada MobiQuitous 2009
  • 2.
    Online Social NetworkSuccess Popular (half billion ww users) Sticky (26m per day)
  • 3.
    OSN Goes MobileAlready top Web destinations on smartphones Unique feature – location GPS-enabled phones Sharing current location Attaching location to user-generated content Outlook LSN >$3.3B revenue by 2013 (ABI) Dodgeball, Loopt, Brightkite, Whrrl Google Latitude, Foursquare
  • 4.
    Brightkite Startup founded2005, Denver CO Angel funding $1M, 03/2008 Private beta, 04/2008 Opened to public, 10/2008 User activity Check in, status update, photo upload All attached with current location Updates through SMS, Email, Web, iPhone… Social graph with mutual connection See your friends’ or local activity streams
  • 5.
  • 6.
    Contributions Data collectionfrom Brightkite 19k users; 1.5m updates Quantitative correlation model for friendship User tags, social graph, location/activity Evaluation using 10m training data and 45d test data Outperformed than Naïve Bayes classifier or J48 decision tree algorithms
  • 7.
    Data Collection BrightkiteWeb APIs 12/9/08-1/9/09: 18,951 active users Back traced to 3/21/08: 1,505,874 updates Profile: age, gender, tags , friends list Social graph: 41,014 nodes and 46,172 links Testing data: next 45 days had 5,098 new links added
  • 8.
  • 9.
  • 10.
    Basic Approach Comingup metrics that Differentiate friends and non-friends Tags, social graph, location, activities Combination of the metrics Training and testing with traces
  • 11.
  • 12.
  • 13.
  • 14.
  • 15.
    Tag Graph 1000most popular tags as the nodes Complete graph Link weight reflects likelihood of two tags shared by friends
  • 16.
  • 17.
  • 18.
  • 19.
  • 20.
    Modeling Accuracy Takeanother 100,000 non-friend pairs Not in training data Plus the newly added 5,098 friend pairs Sort the prediction values
  • 21.
  • 22.
  • 23.
    Information Gain Worldwidebuzz: Planetary-scale views on an instant-messaging network. J. Leskovec and E. Horvitz, June 2007.
  • 24.
    Discussions Model stabilityas Brightkite grows Does not require frequent re-calculation On-demand recommendation Heuristics to speed up metric calculation Possible improvement Different metrics, or combination methods “ Private” updates Conjectured to be few, but no proof
  • 25.
    Related Work Industrialsolutions: Facebook, Twitter Technical details unknown OSN structural analysis Aggregated behavior not suitable for individual recommendations Collective filtering User-item vs. user-user
  • 26.
    Conclusion Correlated attributecombination has good friendship recommendation power Interests, social graph, location Location metric is important Gender and age not so much Future work System implementation Real-user action-based evaluation
  • 27.
    Acknowledgement Anonymous reviewersShepherd- Sharad Agarwal Best Paper Award committee