1 / 23

Using Friendship Ties and Family Circles for Link Prediction

Using Friendship Ties and Family Circles for Link Prediction. Elena Zheleva, Lise Getoor, Jennifer Golbeck, Ugur Kuter (SNAKDD 2008). OUTLINE. Introduction Social Network Model Predicting Links in Social Networks A Feature Taxonomy Experimental Evaluation Conclusions and Future Work.

uma
Télécharger la présentation

Using Friendship Ties and Family Circles for Link Prediction

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 Friendship Ties and Family Circlesfor Link Prediction Elena Zheleva, Lise Getoor, Jennifer Golbeck, Ugur Kuter (SNAKDD 2008)

  2. OUTLINE • Introduction • Social Network Model • Predicting Links in Social Networks • A Feature Taxonomy • Experimental Evaluation • Conclusions and Future Work

  3. INTRODUCTION • There is a growing interest in social media and in data mining methods which can be used to analyze, support and enhance the effectiveness and utility of social media sites. • Social network analysis has focused on actors and relationships between them, such as friendships and family. • There has also been much work in community finding, where densely connected groups of actors are clustered together into communities.

  4. INTRODUCTION • This paper investigate the power of combining friendship and affiliation networks. • The approach here is an attempt to bridge approaches based on structural equivalence and community detection. • Structural equivalence: when two actors are similar based on participating in equivalent relationships. • Two nodesare structurally equivalent if they have the same links to allother actors.

  5. INTRODUCTION

  6. INTRODUCTION

  7. SOCIAL NETWORK MODEL • Social networks describe actors and their relationships. • This paper considers friendship relationships and family group memberships. • The relationships here are undirected, unweighted relationships.

  8. SOCIAL NETWORK MODEL • The networks consist of: • actors: a set of actors A = {a1, . . . , an}. • groups: a group of individuals connected through a common affiliation. The affiliations group the actors into sets G = {G1, . . . , Gm}. • The relationships: • friends: F {ai, aj} denotes that ai is friends with aj. • family: M {ai, Gk} denotes that ai is a part of family Gk. • Attribute b of actor ai : ai.b • The set of friends of actor ai : ai.F • The set of family members of actor ai : ai.M

  9. SOCIAL NETWORK MODEL

  10. SOCIAL NETWORK MODEL

  11. PREDICTING LINKS IN SOCIAL NETWORKS • Link prediction is useful for a variety of tasks. • It is a core component of any system for dynamic network modeling — the dynamic model can predict which actors are likely to gain popularity, and which are likely to become central according to various social network metrics.

  12. PREDICTING LINKS IN SOCIAL NETWORKS • Link prediction is challenging for a number of reasons. • When it is posed as a pair-wise classification problem, one of the fundamental challenges is dealing with the large outcome space; if there are n actors, there are n2 possible relations. • In addition, because most social networks are sparsely connected, the prior probability of any link a priori is extremely small.

  13. A FEATURE TAXONOMY • This paper identified three classes of features in these networks that describe characteristics of potential links in the social network: • Descriptive attributes • Structural attributes • Group attributes

  14. A FEATURE TAXONOMYDescriptive attributes • The descriptive attributes are attributes of nodes in the social network that do not consider the link structure of the network. • Actor features: • Breed • Breed category • Single Breed • Purebred • Actor-pair features • Same breed

  15. A FEATURE TAXONOMYStructural features • These features introduced here describe features of network structure. • Actor features: • Number of friends : |ai.F| • Actor-pair features: • Number of common friends : |ai.F ∩ aj.F| • Jaccard coefficient of the friend sets • Density of common friends

  16. A FEATURE TAXONOMYStructural features • Jaccard coefficient of the friend sets: • The Jaccard coefficient is a standard metric for measuring the similarity of two sets. • Density of common friends: • The number of friendship links between the common friends over the number of all possible friendship links in the set. • The density of common friends of two nodes describes the strength in the community of common friends.

  17. A FEATURE TAXONOMYGroup features • Actor features: • Family Size : |ai.M| • Actor-pair features: • Number of friends in the family : The number of friends ai has in the family of aj : |ai.F ∩ aj.M|. • Portion of friends in the family : The ratio between the number of friends that ai has in aj’s family and the size of aj’s family.

  18. EXPERIMENTAL EVALUATIONData description • Data: a random sample of 10,000 pets each from Dogster and Catster, and all 2059 pets registered with Hamsterster. • For Dogster, the sample of 10,000 dogs had around 17,000 links among themselves, and sample from the non-existing links at a 1:10 ratio. • Using the decision tree classifier from Weka. • The accuracy was measured by computing F1 score.

  19. EXPERIMENTAL EVALUATIONLink-prediction results

  20. EXPERIMENTAL EVALUATIONLink-prediction results

  21. EXPERIMENTAL EVALUATION Alternative network overlays • This paper used the alternative network overlays to test whether there was an advantage to keeping the different types of links and the affiliation groups. • Different-link and affiliation overlay • Same-link and no affiliation overlay • Same-link and affiliation overlay

  22. EXPERIMENTAL EVALUATIONLink-prediction results

  23. CONCLUSIONS ANDFUTURE WORK • This research found that overlaying friendship and affiliation networks were very effective for link prediction. • The experiments show that using affiliation information can achieve significantly higher prediction accuracy. • As future work, investigation on the usage of edge weights and thresholds to define strongly connected clusters, and see if it works as well in link prediction as the family groups did here.

More Related