Konstantin Avrachenkov (INRIA) Prithwish Basu (BBN) Giovanni Neglia (INRIA) Bruno Ribeiro (CMU) Don Towsley (UMass Amherst) 1 K. Avrachenkov, P. Basu, G. Neglia, B. Ribeiro*, and D. Towsley, Pay Few, Influence Most: Online Myopic Network Covering, IEEE NetSciCom Workshop 2014 * corresponding author
(c) 2014, Bruno Ribeiro: Voter Boost on Facebook: Apps targeting supporters 1.Ask campaign contributions (volunteer time, money, etc.) 2.Remind users (recruited nodes) & friends to vote 3.Access to friends list 2
(c) 2014, Bruno Ribeiro: 3 covered friend recruited user Problem: Find largest cover given budget B Each recruitment has unit cost
(c) 2014, Bruno Ribeiro: Common solutions: Minimum Dominating Set (MDS) ◦ NO. Dominating Set must be connected Minimum Connected Dominating Set (MCDS) ◦ Dominating Set is connected 4 REAL-WORLD PROBLEM: TOPOLOGY UNKNOWN
(c) 2014, Bruno Ribeiro: Prioritize invitations without friend degree information Online algorithm 5 covered friend recruited user unknown node
(c) 2014, Bruno Ribeiro: Existing approaches & shortcomings MEED & MOD Conclusions 6
(c) 2014, Bruno Ribeiro: Existing approaches & shortcomings MEED & MOD Conclusions 7
(c) 2014, Bruno Ribeiro: 8 BFS explores nodes in order of discovery FIFO queue priority LM N OP G QH J IK FED BC A
(c) 2014, Bruno Ribeiro: Oracle: (Guha and Khuller’ 98) greedy cover w/known topology BFS Problem: you and your friends have many friends in common (transitivity, cluster) 9 Wiki-talk Slashdot Details in the paper
(c) 2014, Bruno Ribeiro: 10 DFS chooses random unvisited neighbor LIFO queue priority Avoids “cluster” overexploration LM N O P G Q H J IK F ED BC A
(c) 2014, Bruno Ribeiro: Oracle: (Guha and Khuller’ 98) greedy cover w/known topology DFS Problem: ◦ First observed nodes are hubs ◦ Hubs go to bottom of LIFO queue 11 Wiki-talk Slashdot Details in the paper
(c) 2014, Bruno Ribeiro: RW chooses random neighbor No cost of “revisiting” node Random queue priority 12 LM N O P G Q H J IK F ED BC A Random Walk (RW) Search
(c) 2014, Bruno Ribeiro: Oracle: (Guha and Khuller’ 98) greedy cover w/known topology RW advantages: ◦ Less “cluster” problem than BFS ◦ Seeks hubs unlike DFS RW Problem: random priority not targeting potential super- hubs 13 Wiki-talk Slashdot Details in the paper
(c) 2014, Bruno Ribeiro: Existing approaches & shortcomings MEED & MOD Conclusions 14
(c) 2014, Bruno Ribeiro: Enron network 15 Mathematical analysis MUST consider finite graph effects Details in Tech Report Avg ex. degree unrecruited Avg ex. degree unrecruited node with 4 recruited friends Avg ex. degree unrecruited node with 2 recruited friends Avg ex. degree unrecruited node with 1 recruited friend Budget spent so far
(c) 2014, Bruno Ribeiro: (Guha and Kuller’98) myopic heuristic 1. Start tree T = {v} 2. Select neighbors of T with max excess degree 3. Add node to T 4. GOTO 2 until budget exhausted MEED heuristic: Replaces “ with max excess degree” by “ with max EXPECTED excess degree” 16 Excess degree (uncovered degree) Assumes known topology Details in the paper
(c) 2014, Bruno Ribeiro: Chooses node with max recruited neighbors MOD heuristic 1.Select unrecruited w/ max recruited neighbors 2.Invite node 3.GOTO 1 until budget is exhausted In some topologies: node max excess degree = node most recruited friends ◦ e.g., (finite!) random power law graphs with α ∊ {1,2} ◦ approx. true for Erdös-Rényi graphs 17 Details in the paper
(c) 2014, Bruno Ribeiro: Oracle: (Guha and Khuller’ 98) greedy cover w/known topology MOD heuristic: closer to Oracle in all tested social networks 18 Slashdot Wiki-talk Details in the paper
(c) 2014, Bruno Ribeiro: Amazon product-product recommendation network 19 Same nodes, same degrees + randomized neighbors Budget Details in the paper (Maiya & Berger- Wolf, KDD’11) concluded DFS best heuristic for most networks?!?
(c) 2014, Bruno Ribeiro: Existing approaches & shortcomings MEED & MOD Conclusions 20
(c) 2014, Bruno Ribeiro: Myopic Pay-to-cover problems: many open problems with real-world applications ◦ Theory must consider finite networks! Our work: Observations in social networks ◦ Theory: Analysis of finite networks ◦ Empirical + why: DFS consistently bad BFS suffers with clustering RW better than BFS MOD better overall Thank you! Tech 21