1 / 22

Using Transactional Information to Predict Link Strength in Online Social Networks

Using Transactional Information to Predict Link Strength in Online Social Networks. Indika Kahanda and Jennifer Neville Purdue University. Online social networks (OSNs). Explosive growth of online communities enables study of social processes and behavior at a larger scale than ever before

roch
Download Presentation

Using Transactional Information to Predict Link Strength in Online Social Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Using Transactional Information to Predict Link Strength in Online Social Networks • Indika Kahanda and Jennifer Neville • Purdue University

  2. Online social networks (OSNs) • Explosive growth of online communities enables study of social processes and behavior at a larger scale than ever before • Facebook: 200 mil active users • MySpace: 125 mil active users • LinkedIn: 40 mil users • User-contributed data is much more extensive than hand-collected networks previously studied in social science

  3. OSNs are larger and more heterogeneous than manually-collected social networks Purdue Facebook Network UNC National Longitudinal Study of Adolescent Health In-School Survey Min degree=1 Median degree=7 Max degree=10 Min degree=1 Median degree=81 Max degree=2173

  4. High median degree implies the presence of many weak, or spurious, friendship links. Conjecture: Strong relationships can be identified automatically from transactional link information

  5. OSNs contain additional information about user interactions Wall communications Photo postings Group membership

  6. Purdue Facebook network • 56061 public users in March 2008 • Undergrads, grad students, faculty, staff, alumni

  7. Information about strong relationships • Top Friends application allows users to nominate some of their friends as “best friends” • This provides us with positive and negative training examples of strong relationships • 4900 Purdue users have Top Friends application visible publicly (9%) • 17,393 Purdue users are nominated as a Top Friend • Max out-degree=40 max in-degree=14

  8. Automatically identifying top friends • Formulate this as a link strength prediction task • For each friend pair (u,v), predict whether they are “top friends” given their attributes, interactions, and network information. • Use supervised learning methods: Logistic regression, naïve Bayes classifiers, and bagged decision tress • Consider features from four different categories: attributesimilarity, topological connectivity, transactional connectivity, and network-transactional connectivity. • Evaluate on data from the public Purdue Facebook network • Use basic attribute information from profile, friendship links, wall postings, picture postings, group memberships, and “top friend” nominations

  9. Related work • Link prediction • Focuses on predicting future links between any (u,v) pair in a network with a single edge type (i.e., friendship) • Previous methods primarily use attribute similarity features (e.g., Taskar et al. ‘03) or topological features of the network (e.g., Liben-Nowell & Kleinberg ‘04) • Adamic and Adar (‘03) used ancillary network information for link prediction but they focused on similarity-based features instead of transactions/interactions • Pruning spurious links • Singh et al. (’05) and Hill et al. (‘07) sample nodes and edges based on structural properties but they do not consider transactional information

  10. (1) Attribute-based features U V Assess attribute similarity between users (e.g., number of matches) Gender: Male Religious: Christian Political: Moderate Gender: Male Religious: Agnostic Political: Conservative (2) Topological features Assess connectivity of users in friendship network (e.g., number of common neighbors) U V Feature types

  11. (3) Transactional features Wall post Assess transactional activity between user pairs (e.g., number of bi-directional posts) U V Photo post Same group (4) Network-transactional features Assess connectivity of users in transaction networks (i.e., moderate transactional activity by interactions with other users) U V Feature types

  12. Methodology • Models • Bagged decision trees, naïve Bayes classifiers, and logistic regression • Experiments • Feature ranking • Feature type comparison • Link type comparison • Overall classification • Performance measure: area under the ROC curve (AUC) • Measures the quality of (probability) rankings produced by the model

  13. Facebook sample • Random sample of 500 users with top friends application • Consider all friends of those 500 users • Top friends  positive training example • Other friends  negative training example • Restrict attention to pairs that have values for 4 common attributes • Final sample consisted of 8766 linked friends with 896 (10.2%) positive examples

  14. Experiment 1: Feature rankings • Compare relative importance of each of the 50 features • Measures: • Information gain • Chi-square statistic • Compute average rank of each feature and look at top 15: • 12 are network-transactional features, 3 are transactional • 12 use wall information, 3 use picture information

  15. Experiment 2: Feature type comparison • Ablation study using features of each type separately • Attribute-based • Topological • Transactional • Network-transactional • Network-transactional features achieve best performance Network-transactional AUC=84% Transactional AUC=74% Topological AUC=75% Attribute-based AUC=50%

  16. Experiment 3: Link type comparison • Ablation study using data from each link type separately (all features) • Wall • Picture • Groups • Friendship • Wall information results in best performance Friends AUC=77% Wall AUC=82% Group AUC=63% Why doesn’t picture information improve performance?… sparsity. 28% of user pairs have 1 wall link4% of user pairs have 1 picture link Picture AUC=62%

  17. Experiment 4: Overall classification results • Uses 50 features, compares performance of three different models • Bagged decision trees achieve best performance • Network-transactional features account for 97% of the performance observed using all features Bagged Decision Trees AUC=87% Naïve Bayes AUC=81% Logistic Regression AUC=82%

  18. Conclusion • Formulated a link strength prediction task to automatically identify stronger relationships among existing friendships. • Compared the utility of attribute-based, topological, transactional, and network-transactional features • Showed that in addition to good accuracy overall, network-transactional features had the largest impact on model performance • Results indicate that transactional events are useful for predicting link strength • However, it is also necessary to consider the transactional events in the context of user behavior within the larger social network

  19. Future work • Exploit temporal aspect of transactions to improve predictions • Address the more general link-strength prediction task by formulating a latent variable model

  20. Thank you! • Indika Kahanda: • ikahanda@purdue.edu • http://web.ics.purdue.edu/~ikahanda/ • Jennifer Neville: • neville@cs.purdue.edu • http://www.cs.purdue.edu/~neville/ Questions?

More Related