Biao Wang 1, Ge Chen 1, Luoyi Fu 1, Li Song 1, Xinbing Wang 1, Xue Liu 2 1 Shanghai Jiao Tong University 2 McGill University DRIMUX: Dynamic Rumor Influence Minimization with User Experience in Social Networks 1
Outline Introduction of social networks Rumor blocking Proposed algorithms Performance Evaluation Conclusion and future work 2
Social Networks 3 Information sharing and diffusion
Social Network Directed graph: G=(V,E) V - Set of vertices, representing users. E - Set of edges, representing relationships (e.g. user 1 follows user 2)
Social Networks 5 Innovation propagation Information sharing Rumor spreading
Rumor Diffusion 6 Viral spreading (large friends circles ) Causing chaos in society (e.g. ISIS terrorism attack)
7 How do we prevent the rumors from further spreading?
Outline Introduction of social networks Rumor blocking Proposed algorithms Performance Evaluation Conclusion and future work 8
Rumor Propagation SI (Susceptible and Infected) model ( with no recovery ) IC (Independent Cascade) model p 12 p 13 p 14 p43p43 p46p46 p47p47 P ij denotes the probability of node j becoming infected by node i (i, j=1,2,…,7.)
Rumor Propagation How to determine p ij ? 10 Global popularity Topic evolution tendency Individual tendency Sending probability Acceptance probability Ising model in Physics
Rumor Blocking Two strategies: Blocking nodes --- Removing all the edges of the selected node Blocking edges --- Removing selected edges
Rumor Blocking Considering real world problem: will users accept being blocked? 12
Rumor Blocking 13 User experience utility function N nodes, each with blockage threshold T th, which is A constant for homogeneous network A constant for homogeneous network A variable for heterogeneous network A variable for heterogeneous network T b (u) is the blockage time of node u Indicating the average blockage tolerance of whole social network
Outline Introduction of social networks Rumor blocking Proposed algorithms Performance Evaluation Conclusion and future work 14
Problem Formulation Goal Minimize the influence of rumor (the number of infected nodes) Constraint of user experience utility Traditional algorithms fail Time critical Stochastic topology 15
Survival Analysis Survival theory Probability of an event occurring within a time period t If the event occurs during t--- “death”; otherwise, “survival” 16 0 t “Death” “Survive” “Death”
Survival Analysis Survival function The probability that a node “survives” In our context User experience utility determines observation time t A node “survives” means not being infected Our goal --- maximizing the likelihood of nodes “surviving” during the observation time 17
Proposed Algorithms Hazard rate Instantaneous occurrence of an event In our context Hazard rate: expectation of propagation probability from precedent infected nodes Coefficient matrix: Indicator matrix of the network 18
Proposed Algorithms 19 Greedy algorithm Each time finding the optimal node to block K iterations
Proposed Algorithms 20 Dynamic blocking algorithm “Incrementally” finding the optimal node to block Each time blocking nodes
Outline Introduction of social networks Rumor blocking Proposed algorithms Performance Evaluation Conclusion and future work 21
Performance Evaluation Datasets Network extracted from the SinaWeibo, with 23 , 086 nodes, and 183 , 549 edges. Classic Greedy: Greedy algorithm based on descendant order of nodes degree and is used as the baseline algorithm. Greedy algorithm based on descendant order of nodes degree and is used as the baseline algorithm. Proposed Greedy: By blocking a node, we can generate a new propagation matrix and reach a new maximum survival likelihood value. By blocking a node, we can generate a new propagation matrix and reach a new maximum survival likelihood value. Dynamic Algorithm: Adjusts to each propagation status, and gradually includes new targeted nodes as long as the cost is within the scope of tolerable user experience. Adjusts to each propagation status, and gradually includes new targeted nodes as long as the cost is within the scope of tolerable user experience. 22
Performance Evaluation 23 Vertical dashed line indicates the starting point of blocking Left: 54 initial rumor seeds; Right: 32 K=64, the total number of nodes to be blocked
Performance Evaluation 24 Different block durations vs. Infection ratio of network Infection ratio stop decreasing with block duration User experience improvement
Outline Introduction of social networks Rumor diffusion model Rumor blocking algorithms Performance Evaluation Conclusion and future work 25
Conclusion & Future work Conclusion Rumor blocking --- a serious problem in social networks User experience --- a realistic issue in social networks Survival theory --- maximum likelihood solution Dynamic algorithm --- more reasonable and adaptable Future work Network topology --- homogeneous & heterogeneous Experiments on more real world large scale datasets 26
THANK YOU! Q&A THANK YOU! Q&A 27