Price of Anarchy for the N-player Competitive Cascade Game with Submodular Activation Functions Xinran He, David Kempe {xinranhe, dkempe}@usc.edu 12/14/2013 Diffusion In Social Network • The adoption of new products can propagate in the social network ο Diffusion in the social network Competitive Diffusion In Social Network • Different products compete for acceptance in a social network. • Competitive Diffusion in the social network Competitive cascade game • Given a social network πΊ = π, πΈ . • The players are N companies, with their products 1, … , π. • The individuals can be in state 1, … , π and 0. • The players simultaneously allocate resources to individuals in the social network in order to seed them as initial adopters of their products. • The adoption of products propagates according to diffusion model. • The goal for each player is to maximize the coverage of his own product. • In this paper, we study the Price of Anarchy of this game. Main contribution The upper bound on the coarse Price of Anarchy is 2 for the N player competitive cascade game under the Goyal/Kearns diffusion model. Improvement over [Goyal/Kearns 2012]: • Improve PoA upper bound from 4 to 2. • Generalize result from 2 player game to N player game. • Simple and clear proof by resorting to valid utility game and general threshold model. Competitive cascade game • Given a social network πΊ = (π, πΈ). • N players, each player is a company with limit budget π΅π . • Strategy vector for players: πΊ = π1 , … , π π . • π π is the set of nodes selected by company π. • π π ≤ π΅π . • Payoff function π π πΊ : • Expected number of people who adopt product π. • Social utility function πΎ πΊ = π π π (π): π=1 • Expected number of people who adopt a product. General adoption model • Seeding stage: • Each company π selects a set of individuals π π . • The initial state of node π£ is inactive if no company selects it. • Otherwise, the node π£ becomes in state π uniformly at random. • Diffusion stage: • Given a fixed update sequence π = π£1 , … , π£β . • Nodes change states with the order in π according to local dynamics. General adoption model: Local Dynamic • Let πΊ = (π1 , … , π π ) be current sets of nodes in state π. • Adoption function: • βπ£π πΊ = Prob{π£ adopts product π} • Total activation probability: • π»π£ πΊ = π π β π=1 π£ πΊ • A still inactive node π£ changes into states π with probability βπ£π (πΊ), and remains inactive with probability 1 − π»π£ (πΊ). General adoption model: Example πΊπ© Diffusion stage G D B F C A Seeding stage D C D F C END E ππ Useful properties Additivity of total activation probability π ), activation function π π is monotone. π»π£ πΊ = ππ£ (∪π π π£ π=1 Submodularity of activation function: ππ£ π ∪ π£ −Prob{ ππ£ π ≥} ππ£ π≤ ∪ π£ Prob{ − ππ£ π , ∀ π ⊆π } Competitiveness of adoption function: π , π −π ⊆ π −π , π −π =∪ π βπ£π π ≥ βπ£π π Prob{ , ∀π π ⊆ π π = π≠iProb{ ? } ? } Main results Theorem: Assume the following conditions hold: 1. 2. 3. For every node π£, the total activation probability π»π£ π is additive. For every node π£, the activation function ππ£ π is submodular. For every player π and node π£, the adoption function βπ£π π is competitive. Then, the upper bound on the coarse PoA is 2 in the competitive cascade game. Improvement over [Goyal/Kearns 2012]: • Improve PoA upper bound from 4 to 2. • Generalize result from 2 player game to N player game. Proof roadmap Submodularity of social utility function πΎ(⋅) By reduction to general threshold model Set Game ππ πΊπ ≥ πΎ πΊπ − πΎ(πΊπ−π , ∅π ) By global competitiveness ππ π π0 ≤ πΎ(π0 ) By definition of social utility function Valid utility game [Vetta 2002] [Roughgarden 2009] PoA bounds Proof roadmap Submodularity of social utility function πΎ(⋅) By reduction to general threshold model Set Game ππ πΊπ ≥ πΎ πΊπ − πΎ(πΊπ−π , ∅π ) By global competitiveness ππ π π0 ≤ πΎ(π0 ) By definition of social utility function By definition. Valid utility game [Vetta 2002] [Roughgarden 2009] PoA bounds Proof roadmap Submodularity of social utility function πΎ(⋅) By reduction to general threshold model Set Game ππ πΊπ ≥ πΎ πΊπ − πΎ(πΊπ−π , ∅π ) By global competitiveness ππ π π0 ≤ πΎ(π0 ) By definition of social utility function Valid utility game [Vetta 2002] [Roughgarden 2002] PoA bounds Submodular πΎ(⋅): General Threshold model • General Threshold (GT) Model [KKT 03] • • • • Each node has a threshold ππ£ uniform in [0,1] Each node has an activation function, ππ£ π , π is the set of activated nodes. A node becomes active if and only if ππ£ π ≥ ππ£ . π(π) is expected number of activated nodes at the end of the process. Theorem [Mossel/Roch 2007]: Under the general threshold model with monotone and submodular ππ£ (S) , σ(S) is monotone and submodular. Submodular πΎ(⋅): reduction to GT model ππ ππ ππ ππ ππ ππ ππ ππ Update sequence: ππ ππ ππ π πβ πβ πβ πβ ππ π π … π π π Active Inactive Proof roadmap Submodularity of social utility function πΎ(⋅) By reduction to general threshold model Set Game ππ πΊπ ≥ πΎ πΊπ − πΎ(πΊπ−π , ∅π ) By global competitiveness ππ π π0 ≤ πΎ(π0 ) By definition of social utility function Valid utility game [Vetta 2002] [Roughgarden 2009] PoA bounds Proof of π π πΊπ ≥ πΎ πΊπ − −π π πΎ(πΊπ , ∅ ) π) • Global competitiveness: π π πΊπ ≤ π π (πΊ−π , ∅ π • Similar to Lemma 1 in [Goyal/Kearns 2012] π • Couple two process ππ‘ with πΊπ and ππ‘ with (πΊ−π , ∅ ). π • By induction, ππ‘π ⊇ ππ‘π , ππ‘0 ⊆ ππ‘0 ⇒ ππ‘−π ⊆ ππ‘−π πΏπ βπ£π (πΏπ ) βπ£−π (πΏπ ) 1 − π»π£ (πΏπ ) βπ£π (ππ ) βπ£−π (ππ ) 1 − π»π£ (ππ ) ππ Proof: wrap up Lemma: social utility function πΎ(⋅) is submodular, if π»π£ πΊ is additive and ππ£ (π) is submodular. π , Lemma: π π πΊπ ≥ πΎ πΊπ − πΎ πΊ−π , ∅ π if π»π£ πΊ is additive and βπ£ (π) is competitive. The competitive cascade game is a valid utility game The pure PoA is bounded by 2 [Vetta 2002] The coarse PoA is bounded by 2 [Roughgarden 2009] Conclusion • Improvement over [Goyal/Kearns 2012]: • Improve PoA upper bound from 4 to 2. • Generalize from 2 players to N players. • Generalize from pure PoA to coarse PoA. • With a much simpler and clear proof. • Further extensions: • Strategy as multiset: πΌπ£π π ≤ πΎπ • Budget limit on nodes: π πΌ π£ π=1 π£ • Different node weight ππ£ , π π π = π[ π£∈πβπ ππ£ ] Future work • Open question • What is the PoA upper bound for competitive cascade game without submodularity of activation function? • Upper bound 4 with additive total activation probability and competitive adoption function for 2 player games. [Goyal/Kearns 2012] • Lower bound 2 by simple example. • Results on cascade without submodularity • Influence maximization: • Single product: submodularity -> 1 − 1/π. [KKT 2003] • Competitive cascade game