A Note on Competitive Diusion Through Social Networks

Noga Alon Michal Feldman Ariel D. Procaccia Moshe Tennenholtz

Abstract We introduce a game-theoretic model of diusion of technologies, advertisements, or inuence through a social network. The novelty in our model is that the players are interested parties outside the network. We study the relation between the diameter of the network and the existence of pure Nash equilibria in the game. In particular, we show that if the diameter is at most two then an equilibrium exists and can be found in polynomial time, whereas if the diameter is greater than two then an equilibrium is not guaranteed to exist.

Introduction

Social networks such as Facebook and Twitter are modern focal points of human interaction. The pursuit of insights into the nature of this interaction calls for a game-theoretic analysis. Indeed, a number of papers (see, e.g., [5]) investigate variations on the following setting. The social network is represented by an undirected graph, where the vertices are users and edges connect users who are in a social relationship. Suppose, for example, that there are several competing applications, e.g., voice over IP systems, that are not interoperable. The users play a coordination game, where if two neighbors adopt the same system they get some reward that is based on the inherent quality of the system. The goal is to study the diusion of technologies through the social network. The point of view here is completely decentralized, and the players in the game are the users of the social network. We propose a dierent, global point of view regarding the incentives that govern the diusion process. Suppose we have several rms that would like to advertise competing products via viral marketing. Each rm initially targets a small subset of users, in the hope that the rumor about its product would spread throughout the network. However, a user that adopts one product is reluctant to adopt another, hence the campaign of one rm negatively aects the success of another rms campaign. To the best of our knowledge our model is the rst game-theoretic model to deal with
Microsoft Israel R&D Center, 13 Shenkar Street, Herzeliya 46725, Israel, and Schools of Mathematics and Computer Science, Tel Aviv University, Tel Aviv, 69978, Israel, Email: [email protected]. Research supported in part by a USA Israeli BSF grant, by a grant from the Israel Science Foundation, by an ERC advanced grant and by the Hermann Minkowski Minerva Center for Geometry at Tel Aviv University. Microsoft Israel R&D Center, 13 Shenkar Street, Herzeliya 46725, Israel, and School of Business Administration and Center for the Study of Rationality, The Hebrew University of Jerusalem, Jerusalem 91904, Israel. Email: [email protected] School of Engineering and Applied Sciences, Harvard University, 33 Oxford Street, Cambridge, MA 02138, United States. Email: [email protected]. This work was done while the author was at Microsoft Israel R&D Center. Microsoft Israel R&D Center, 13 Shenkar Street, Herzeliya 46725, Israel, and Technion, IIT, Haifa 32000, Israel. Email: [email protected]

2 1 1 1 2

2 1 1 2

(a) Time 1.

(b) Time 2.

(c) Time 3, the process terminates.

Figure 1: An illustration of the diusion process, with N = {1, 2}. the incentives of interested parties outside the social network. Note that some previous papers did consider the problem of choosing an inuential set of users as an optimization problem (see, e.g., [6]), but not in a competitive game-theoretic setting. Other papers, which deal with Voronoi games on graphs, provide a game-theoretic study of a facility location problem that does not involve a diusion process, where rather each vertex is assigned to the closest agent and the utility of an agent is the number of vertices assigned to it (see, e.g., [3, 7]). The model. Let G = V, E be an undirected graph. Furthermore, let N = {1, . . . , n} be the set of agents (the interested parties). The diusion process unfolds as follows. There are n + 2 colors: a color for each agent i N , as well as two additional colors: white and gray. Initially, at time 1, some of the vertices are colored in the colors of N , while the others are white. At time t + 1 each white vertex that has neighbors colored in color i, but does not have neighbors colored in color j for any j N \ {i}, is colored in color i. A white vertex that has two neighbors colored by two distinct colors i, j N is colored gray. In other words, we assume that if two agents compete for a user at the same time they cancel out and the user is removed from the game. The process continues until it reaches a xed point, that is, all the remaining white vertices are unreachable due to gray vertices. See Figure 1 for an illustration of the diusion process. A game = G, N is induced by a graph G, representing the underlying social network, and the set of agents N . The strategy space of each agent is the set of vertices V in the graph, that is, each agent i selects a single node that is colored in color i at time 1. Note that if two or more agents select the same vertex at time 1 then that vertex becomes gray. A strategy prole is a vector x = x1 , . . . , xn V n , where xi V is the initial vertex selected by agent i. We also denote xi = x1 , . . . , xi1 , xi+1 , . . . , xn . Given a strategy prole x V n , the utility of agent i N , denoted Ui (x), is the number of nodes that are colored in color i when the diusion process terminates. For instance, in the example given in Figure 1 the utility of each of the agents is two. A strategy prole x is a (pure strategy) Nash equilibrium of the game if an agent cannot benet from unilaterally deviating to a dierent strategy, i.e., for every i N and xi V it holds that Ui (xi , xi ) Ui (x). Our results. Given a graph G and u, v V , let d(u, v ) be the length of the shortest path between u and v (in terms of the number of edges). The diameter of the graph, denoted D(G), is the maximum distance between a pair of vertices, that is, D(G) = maxu,vV d(u, v ). Our investigation focuses on the relation between the diameter of the graph and the existence of Nash equilibria in the induced diusion game. Indeed, if we can nd a Nash equilibrium then we can often predict the behavior of the agents and the outcome of this competitive diusion process, or, alternatively, advise the agents how to play. Our rst theorem is the following. 2

Theorem 2.1. Every game = G, N where D(G) 2 admits a Nash equilibrium. Furthermore, an equilibrium can be found in polynomial time. Note that a random graph on n labeled vertices where each edge appears with probability p, usually denoted G(n, p), has diameter at most two with high probability whenever p (c ln n)/n for c > 2 (see, e.g., [2] for more details about the diameter of random graphs). In particular (by taking p = 1/2) almost all graphs over n vertices have diameter at most two. Finally, social networks typically have a very small diameter. Therefore, it can be argued that assuming a diameter of two is not very restrictive. It is now natural to ask whether the existence of Nash equilibria can also be guaranteed for diameters larger than two. It is not too dicult to construct a graph with diameter four that does not admit an equilibrium. Our second theorem gives a negative answer even with respect to diameter three. Theorem 2.2. Let N = {1, 2}. There exists a graph G with D(G) = 3 such that the game = G, N does not admit a Nash equilibrium. The construction in the proof of Theorem 2.2 can easily be extended to a larger number of agents or to any (nite or innite) diameter greater than three. Discussion. In order to facilitate the game-theoretic analysis we consider a very simple model of diusion. In particular, conicts are deterministically resolved by introducing gray vertices, and each agent initially selects just one vertex. Richer (probabilistic) models of diusion through a social network exist in the literature, e.g., [6, 4]. On the other hand, the assumption of discrete time steps is quite common. Theorem 2.1 implies that with high probability a random graph (even a relatively sparse one) induces a game that admits a Nash equilibrium. However, social networks are normally not completely random, but rather often exhibit structure. Ideally one would be able to extend our result by showing that under a convincing random graph model of social networks (see, e.g., [1, 9]) the induced game admits a Nash equilibrium with high probability.

Proofs

We begin by proving Theorem 2.1; we subsequently discuss some implications of the proof. Theorem 2.1. Every game = G, N where D(G) 2 admits a Nash equilibrium. Furthermore, an equilibrium can be found in polynomial time. Proof. If D(G) 1 then the graph is a clique and the theorem follows trivially. Therefore, we may assume that D(G) = 2. Given a prole x V N , let P (x) = |{ i, j : d(xi , xj ) = 1}|, that is, the number of pairs with distance one from each other. Furthermore, denote the neighborhood of vertex u V by Nu = {v : d(u, v ) 1}, and let N (x) = n i=1 Nxi . Consider the potential function (x) = |N (x)| n + P (x) . It is sucient to show that for every x V n , i N , and xi V , Ui (xi , xi ) > Ui (x) (xi , xi ) > (x) . 3 (1)

Indeed, given Equation (1) it clearly holds that any strategy prole x V n that maximizes (x) must be a Nash equilibrium. Moreover, in order to nd one such prole we may start from some preference prole, and in each step attempt to nd a protable deviation for one of the agents. We terminate if there is no such deviation (which, by denition, means that we have found a Nash equilibrium). This algorithm terminates after a polynomial number of steps since (x) is bounded from above by n|V | + n2 for every x, and by Equation (1) every protable deviation by an agent increases the value of the potential function by at least one. We turn to proving Equation (1). If the diameter of the graph is two then vertices can only be colored by an agent i N at time 1 or 2. Specically, the vertices colored by agent i are roughly the vertices in the neighborhood of xi that are not neighbors of xj for some j N \ {i} (since these vertices are either gray or colored by j ). Formally, dene Ai = {x : j N \ {i} s.t. d(xi , xj ) = 1} . Assuming that xi = xj for all i = j , the utility of agent i under the strategy prole x V n is Ui (x) = |Nxi | |
j =i

(Nxi Nxj )| + Ai (x) ,

where Ai is the indicator function that returns 1 if x Ai and 0 otherwise. The rightmost term is required since even if xi is a neighbor of some xj , it is still colored by agent i at time 1, but is nevertheless included in the middle term. Now, suppose Ui (xi , xi ) > Ui (x). It follows that |Nxi | |
j =i

(Nxi Nxj )| + Ai (xi , xi ) > |Nxi | |

(Nxi Nxj )| + Ai (x) .

(2)

Since Ai is a Boolean function, this implies that |Nxi | |

(Nxi Nxj )| |Nxi | |

(Nxi Nxj )| .

(3)

We distinguish between two cases. If Equation (3) holds as a strict inequality then |
j =i

Nxj Nxi | = |
j =i

Nxj | + |Nxi | |
j =i

(Nxi Nxj )| > |

Nxj | + |Nxi | |
j =i

(Nxi Nxj )|

=|
j =i

Nxj Nxi | ,

which implies that |N (xi , xi )| |N (x)| + 1. In addition, a deviation of a single agent can decrease the number of adjacent pairs of agents by at most n 1, i.e., P (xi , xi ) > P (x) n. We conclude that (xi , xi ) = |N (xi , xi )|n+P (xi , xi ) |N (x)|n+n+P (xi , xi ) > |N (x)|n+n+P (x)n = (x) . Otherwise, Equation (3) holds as an equality, and hence |N (xi , xi )| = |N (x)|. It then follows from Equation (2) that Ai (xi , xi ) > Ai (x). That is, agent i has no neighbors among xi under xi but has at least one neighbor under xi . Thus the number of neighbors of agent i increases and 4

the number of neighbors of agents j N \ {i} does not decrease, i.e., P (xi , xi ) > P (x). We conclude that (xi , xi ) = |N (xi , xi )|n + P (xi , xi ) = |N (x)|n + P (xi , xi ) > |N (x)|n + P (x) = (x) . This establishes Equation (1), and hence completes the proof of the theorem. What the proof of Theorem 2.1 essentially shows is that when the diameter of the graph is two the diusion game is a potential game [8]; specically, a function that satises (1) is known as a generalized ordinal potential function. Potential games have the property that better response dynamics converge to a Nash equilibrium; in other words, if at every stage the agents simply behave myopically, that is, some agent deviates to a more protable strategy, then they will eventually reach an equilibrium. We are now ready to prove our second theorem. Theorem 2.2. Let N = {1, 2}. There exists a graph G with D(G) = 3 such that the game = G, N does not admit a Nash equilibrium. Proof. We rst give our construction, then establish that it has diameter three and that it does not admit a Nash equilibrium. The construction. Let G = V, E be dened as follows. The vertices of the graph are V = {v1 , . . . , v6 } C1 C2 C3 , where for i = 1, 2, 3, Ci = Ci1 Ci5 . Each Cij contains ten vertices, that is, |V | = 156. The edges of the graph are dened as follows. Each Ci , for i = 1, 2, 3, is a clique. There is an edge v1 , u for every u C11 C12 C13 C21 C22 C23 ; an edge v2 , u for every u C11 C14 C15 C21 C24 C25 ; an edge v3 , u for every u C11 C12 C14 C31 C32 C33 ; an edge v4 , u for every u C11 C13 C15 C31 C34 C35 ; an edge v5 , u for every u C21 C22 C24 C31 C32 C34 ; an edge v6 , u for every u C21 C23 C25 C31 C33 C35 . An illustration of the graph G is given as Figure 2. We refer to the vertices v1 , . . . , v6 as hubs ; we say that v1 and v2 are parallel hubs, and so are v3 and v4 , v5 and v6 . If the hub vi is connected by an edge to some of the vertices of clique Cj , we say that vi is adjacent to Cj ; for instance, v1 and C1 are adjacent, whereas v1 and C3 are not. The construction possesses the following important properties: 1. Let vi and vj be two parallel hubs that are adjacent to a clique Ck . Then (Nvi \ Nvj ) Ck contains exactly two of the sets Ckl , l = 1, . . . , 5. 2. Let vi and vj be two nonparallel hubs that are adjacent to a clique Ck . Then (Nvi \ Nvj ) Ck contains exactly one of the sets Ckl , l = 1, . . . , 5. Note that the construction is essentially symmetric with respect to the hubs. G has diameter 3. Using Figure 2, it is easy to verify that G has diameter 3. For example, a path from v1 to u C32 is given by v1 w v3 u, where w C11 . A path from u C13 to w C24 is given by u v1 x w, where x C21 . G does not admit a Nash equilibrium. We consider strategy proles x1 , x2 V 2 for the two agents. The symmetries of our construction allow us to restrict our attention to six cases. 5

v1

v2

C11

C12

C13

C14

C15

C21

C22

C23

C24

C25

v3

v4

v5

v6

C31

C32

C33

C34

C35

Figure 2: The construction of the proof of Theorem 2.2. The cliques C1 , C2 , C3 are outlined by dashed ellipses, and the edges inside the cliques are not shown. An edge between vi and Cjk implies that vi is connected to all the vertices u Cjk . Case 1: x1 = v1 , x2 C1 (hub and adjacent clique). Agent 1 colors some of the vertices of C2 and some hubs, that is, U1 (x1 , x2 ) < 60. By deviating to x1 = v5 , agent 1 colors C21 , C22 , C24 , C31 , C32 , C34 , i.e., U1 (x1 , x2 ) 60. Case 2: x1 = v1 , x2 C3 (hub and nonadjacent clique). Agent 2 colors the vertices of C3 and some hubs, hence U2 (x1 , x2 ) < 60. By deviating to x2 C11 , agent 2 colors C14 , C15 , and C3 , thus U2 (x1 , x2 ) 70. Case 3: x1 = v1 , x2 = v3 (nonparallel hubs). Agent 1 colors C13 , C2 , and some hubs, therefore U1 (x1 , x2 ) < 70. By deviating to x1 C11 , agent 1 colors C13 , C15 , C2 , so U1 (x1 , x2 ) 70. Case 4: x1 = v1 , x2 = v2 (parallel hubs). Agent 1 colors C12 , C13 , C22 , C23 , and some hubs (v3 , . . . , v6 are gray and C3 remains white), hence U1 (x1 , x2 ) < 50. By deviating to x1 C3 , agent 1 can guarantee a utility of at least 50 (since it colors C3 ). Case 5: x1 C1 , x2 C3 (dierent cliques). If x1 / C11 , x2 / C31 , then agent 1 can benet by deviating to x1 C11 , since then it colors both v1 and v2 at time 2 (rather than just one of them), and colors twenty vertices of C2 at time 3 (rather than ten). Hence we can assume without loss of generality that x1 C11 . In that case, agent 2 colors C3 and some hubs, therefore U2 (x1 , x2 ) < 60. By deviating to x2 = v5 , agent 2 colors at least C21 , C22 , C24 , C31 , C32 , C34 , hence U2 (x1 , x2 ) 60. Case 6: x1 C1 , x2 C1 (same clique). Since C1 \ {x1 , x2 } is gray, there are at most 108 vertices that are not gray, therefore it must hold that either U1 (x1 , x2 ) < 60 or U2 (x1 , x2 ) < 60. By deviating to v5 an agent can guarantee a utility of at least 60.

References
[1] A. L. Barab asi and R. Albert. Emergence of scaling in random networks. Science, 286:509512, 1999. [2] B. Bollob as. Random Graphs. Cambridge University Press, 2nd edition, 2001. [3] C. D urr and N. K. Thang. Nash equilibria in Voronoi games on graphs. In Proceedings of the 15th Annual European Symposium on Algorithms (ESA), pages 1728, 2007. 6

[4] R. Durrett. Random Graph Dynamics. Cambridge University Press, 2006. [5] N. Immorlica, J. M. Kleinberg, M. Mahdian, and T. Wexler. The role of compatibility in the diusion of technologies through social networks. In Proceedings of the 8th ACM Conference on Electronic Commerce (EC), pages 7583, 2007. Tardos. Maximizing the spread of inuence through a social [6] D. Kempe, J. M. Kleinberg, and E. network. In Proceedings of the 9th International Conference on Knowledge Discovery and Data Mining (KDD), pages 137146, 2003. [7] M. Mavronicolas, B. Monien, V. G. Papadopoulou, and F. Schoppmann. Voronoi games on cycle graphs. In Proceedings of the 33rd International Symposium on Mathematical Foundations of Computer Science (MFCS), pages 503514, 2008. [8] D. Monderer and L. S. Shapley. Potential games. Games and Economic Behavior, 14:124143, 1996. [9] M. E. J. Newman, D. J. Watts, and S. H. Strogatz. Random graph models of social networks. Proceedings of the National Academy of Sciences, 99:25662572, 2002.