Understanding Community Effects on Information Diffusion!

Transcription

1 Understanding Community Effects on Information Diffusion! Shuyang Lin Qingbo Hu Guan Wang Philip S. Yu University of Illinois at Chicago Presented by Hong- Han Shuai NaConal Taiwan University

2 2 Picture credit: Tanja Scherm, Flickr

3 Information diffusion in social network!

4 Information diffusion in social network!

5 Viral marketing! advercsers users

6 Influence maximization problem! x y w z u Given an informacon diffusion model, find a seed set of k accve users such that the expected number of accve users is maximized. Example: k=1, {z} v s t k=2, {z, s} Generally, a very hard problem (NP- hard for almost all popular informacon diffusion models). 6

7 IC model and influence maximization! Independent Cascade model Whenever a user becomes accve, he has an independent chance to make each of his neighbors accve. q y 0.2 z x v 0.3 Influence maximizacon problem on the IC model Standard algorithm: greedy, select the node that maximize increase of influence in each step Speed- up algorithms: pruning Use random sampling to calculate that. Very Cme- consuming. 0.5 w r 0.5 u 7

8 Motivation! CommuniCes ubiquitously exist in social networks. Community structure of a social network can be considered as a summarizacon of it. Such summarizacon may help us come up with some efficient approximate algorithms for influence maximizacon problem. 8

9 Community effects on information diffusion! Goal: Understand the effects of communices on informacon diffusion by empirical studies on real- world data. Methodology: 1. Use classic community deteccon algorithms to detect communices in a social network. 2. Study quescons such as: Whether users in the same communices have similar behaviors? Whether users in the same communices are influenced by similar influencers? Datasets: Foursquare, DBLP Community deteccon methods Fast greedy (FG), Infomap (IM) 9

10 a 1 a 2 Action homophily of communities! - Whether users in the same communices have similar accons? 1. Construct accon vector a i for each user v i 1 2 j a ij = 1, if v i takes accon in the j- th cascade. a ij = 0, otherwise. a i a i - accon vector of user v i 2. Calculate cosine similarity between a i and a j for 3 cases: v i and v j are friends in the same community v i and v j are friends in different communices v i and v j is an arbitrary pair of users 10

11 Distribution of action similarity! 11

12 Role-based homophily of communities! - - Whether users in the same communices have similar influencers? Whether users in the same communices influence similar users? 1. Construct matrix of influence S j S ij influence from v i to v j i s i* influencer feature vector of v i s *j influencee feature vector of v j 2. Calculate cosine similarity between influencer feature vectors or influencee feature vectors for three cases as before. 12

13 Distribution of influencer feature vector similarity! 13

14 Distribution of influencee feature vector similarity! 14

15 Observations! We observed accon- based homophily, influencer- role- based homophily, and influencee- role- based homophily from real- world datasets. Role- based homophily is much stronger than the accon- based homophily. Influencee- role- based homophily is more significant than the influencer- role- based homophily 15

16 Community-based Fast Influence Model! An efficient approximate algorithm for influence maximizacon using influencee- role- based homophily. v 1 v 2 Community detection v 2 v 1 v 3 G v 4 v 3 v 4 G b v 1 v 4 v 2 v 3 v 1 v 4 v 2 v 3 in v in 1 v 2 in in v v 3 4 Influence decoupling v 1 in v 2 inv 3 in v 4 in 16 Influence aggregation

17 Influence decoupling! G v 1 v 2 v 4 G b v 2 v 1 v 4 in v in 1 v 2 in v 4 v 3 v 3 v 3 in Separate the roles of users as influencers and influencees. Transfer G to a biparcte graph G b Transfer each node v i in the original graph to two nodes. 17

18 Community detection and influence aggregation! AgglomeraCve clustering method: Aggregate influence Merge Community Repeat two steps alternately uncl the similarity between any two communices is smaller than a threshold θ. 18

19 Influence maximization! Select k nodes that maximize the total influence Select the node that maximize the increase of total influence z Deduct the number of nodes that are influenced by the selected node 19

20 Experiment! Datasets: Foursquare DBLP Algorithms: CFIGreedy (proposed method) ICGreedy (Influence maximizacon on IC with CELF++ opcmizacon) Degree DegreeDiscount Random 20

21 Effectiveness results! 21

22 Efficiency results! 22

23 Thank you!!