Marketing in a random network

Marketing in a random network

Hamed Amini INRIA, ENS Paris and Imperial College London    Moez Draief INRIA, ENS Paris and Imperial College London    Marc Lelarge INRIA, ENS Paris and Imperial College London
Abstract

Viral marketing takes advantage of preexisting social networks among customers to achieve large changes in behaviour. Models of influence spread have been studied in a number of domains, including the effect of “word of mouth” in the promotion of new products or the diffusion of technologies. A social network can be represented by a graph where the nodes are individuals and the edges indicate a form of social relationship. The flow of influence through this network can be thought of as an increasing process of active nodes: as individuals become aware of new technologies, they have the potential to pass them on to their neighbours. The goal of marketing is to trigger a large cascade of adoptions. In this paper, we develop a mathematical model that allows to analyze the dynamics of the cascading sequence of nodes switching to the new technology. To this end we describe a continuous-time and a discrete-time models and analyse the proportion of nodes that adopt the new technology over time.

Keywords:
models of contagion, random graphs

1 Introduction

With consumers showing increasing resistance to traditional forms of advertising, marketers have turned to alternate strategies like viral marketing. Viral marketing exploits existing social networks by encouraging customers to share product information with their friends. Social networks are graphs in which nodes represent individuals and edges represent relations between them. To illustrate viral marketing, consider a company that wishes to promote its new instant messenger (IM) system [7]. A promising way would be through popular social network such as Myspace: by convincing several persons to adopt the new IM system, the company can obtain an effective marketing campaign and diffuse the new system over the network.

If we assume that “convincing” a person to “spread” the new technology costs money, then a natural problem is to detect the influential members of the network who can trigger a cascade of influence in the most effective way [2], [5]. In this work, we consider a slightly different problem: the marketer has no knowledge of the social network. Hence he will not be able to detect the most influential individuals and his only solution is to “convince” a fraction of the total population. However, the marketer can still use the structure of the underlying network by targeting the neighbours of the adopters. There are a number of incentive programs around this idea: each time an individual chooses the new technology, he is given the opportunity to send e-mail to friends with a special offer; if the friend goes on to buy it, each of the individuals receives a small cash bonus.

In this paper, we develop a mathematical model that allows to analyze the dynamics of the cascading sequence of nodes switching to the new technology. To this end we describe a continuous-time and a discrete-time models and analyze the proportion of nodes that adopt the new technology over time. In the continuous setting we derive a general bound for the proportion of new adopters in terms global graph properties, namely the spectral radius and the minimum degree. In the discrete setting we show that the proportion of new adapters is the solution of a fixed point equation. To this end we examine the case of regular trees, and prove that our approach carries over to random regular graphs. We extend our model to the general threshold model [5] and to sparse random graphs. We conclude by presenting a framework that enables the control of the marketing policy and discuss other possible applications.

2 Model

We consider a set of agents represented by an undirected graph structure accounting for their interaction. For , we write if and we say that agents and are neighbours. As in [11], we consider binary models where each agent may choose between two possible strategies that we denote by and . Let us introduce a game-theoretic diffusion model proposed by Morris [9]: Whenever two neighbours in the graph opt for strategy they receive a payoff , if they both choose they receive a payoff , and they receive nothing if they choose opposite strategies. The payoff of an agent corresponds to the sum of its payoffs with each of its neighbours.

Initially all nodes play except for a small number of nodes that are forced to adopt strategy . The nodes that started with strategy will subsequently apply best-response updates. More precisely, these nodes will be repeatedly applying the following rule: switch to if enough of your neighbours have already adopted . There can be a cascading sequence of nodes switching to such that a network-wide equilibrium is reached in the limit. This equilibrium may involve uniformity with all nodes adopting or it may involve coexistence, with the nodes partitioned into a set adopting and a set sticking to .

The state of agent is represented by ; if player plays strategy and otherwise. Hence is the number of neighbours of playing strategy and is the number of neighbors of playing strategy .

We now describe the economic model for the agents. Recall that the payoff for a edge is , for a edge is and for a edge is . We assume that if an agent chooses , his payoff is just the sum of the payoffs obtained on each of his incident edges but if he chooses , his payoff is the sum of these payoffs increased by an amount plus a bonus of . Now the total payoff for an agent is given by

(1)

We consider that and are fixed and correspond to the level of performance of the technologies and .

By (1), we have iff

(2)

with and is the degree (number of neighbours) of . We now explain the dynamics of our model for the spread of strategy in the network as time evolves. We consider a fixed network (not evolving in time) and let all agents play for . At time , some agents are forced to strategy . These agents will always play strategy , hence the dynamics described below does not apply to these initially forced agents. We encode the initial population forced to strategy by a vector , where if agent is forced to and otherwise. We will assume that the vector is a sequence of i.i.d. Bernoulli random variables with parameter .

3 Continuous-time dynamic

We first consider the following continuous version of the contagion model. Assume that each non infected node updates its state at rate and it holds on to strategy if (2) is satisfied and switches to B if . The state at time is represented by a vector . Denote by the adjacency matrix of the graph and let the spectral radius of , namely, its largest eigenvalue and by the minimum degree of graph . In addition we will assume that the graph is connected so that and has multiplicity one. Therefore we have, , for all , and

Note that We now consider the continuous time Markov process , with , and transition rate:

standard coupling arguments yield for all , where denotes that stochastically dominates . This implies that . Moreover, the transition rates of the process are such that

Hence

(3)

Using Cauchy-Schwartz inequality, we obtain that . Combining this with (3), we have that

Theorem 3.1

Let be the proportion of nodes that opted for strategy by time . Then

Moreover if the is a regular graph with degree , then, using the spectral decomposition of the matrix , we have that

The above result states that the number of nodes that have adopted increases at most exponentially in time and that the speed is given by . Similar results have been found in [3] in the case of the Susceptible-Infected-Susceptible (SIS) epidemic.

As a matter of example for Erdös-Rényi graphs with parameters and in the regime , Theorem 3.1 yields with high probability.

In the next section, we describe a discrete-time version of our contagion model for which we derive more accurate results for the proportion of -adopters and illustrate the coexistence of the two strategies.

4 Discrete-time dynamic

The state of the network at time is described by the vector , . We have and . Then at each time step , each agent applies the best-response update: if then he chooses and if not then he chooses . It is readily seen that

(4)

4.1 Diffusion process on the infinite regular tree

Let be an infinite -regular tree with nodes , with a fixed root . For a node , we denote by the generation of , i.e. the length of the minimal path from to . Also we denote if belongs to the children of , i.e. and is on the minimal path from to . For an edge with , we denote by the sub-tree of with root obtained by the deletion of edge from .

For a given vector , we say that node is infected from if the node switches to in with the same vector for and the strategy A for . We denote by the corresponding indicator function with value if is infected from at time and otherwise.

Proposition 1

We have

(5)

The representation (5) is crucial to our analysis. In fact, thanks to the tree structure, the random variables are independent of each other and identically distributed. More precisely, a simple induction shows that (4) becomes, for :

(6)

Note that (6) allows to compute all the recursively, starting with . It is then easy to compute their distribution from (6). We summarize this result in the next proposition.

Proposition 2

For fixed, the sequence is a sequence of i.i.d. Bernoulli random variables with parameter given by and, for ,

where . corresponds to the binomial distribution with parameters and .

Combining Propositions 1 and 2, we obtain that

Theorem 4.1

is a Bernoulli random variable with parameter given by

(7)

Moreover let the smallest solution of the following fixed point equation

(8)

Suppose . There exists such that for all , the fixed point equation (8) has a unique solution and for all it has three solutions .

4.2 Random regular graphs

We now come back to the process on , a random regular graph, satisfying (4). Given , let be the set of vertices of that are at a distance at most from . A depth- -regular tree is the restriction of to nodes with . A simple induction on shows that is determined by the . Using the following convergence [4]: for any fixed , we have as tends to infinity , we have as tends to infinity. Therefore the process defined on the tree in Section 4.1 is a good approximation of the real process. Hence,

Proposition 3

For any fixed , we have

(9)

where is defined in (7).

Let be the proportion of agents choosing at time : . We have as ,

(10)

The final proportion of agents choosing is .

Proposition 4

We have

(11)

in particular, for , we have

(12)

The interchange of limits in and needs a proper mathematical proof. This has been done in [1] and our statement follows from their Theorem 1. For -regular graphs, bootstrap percolation is equivalent to our model. It is noticed in [1] that the critical value on the -regular random graph turns out to be the same as that on the -tree, although the proof goes along a quite different route.

4.3 Extensions: random networks and linear threshold model

In this section, we show how our approach extends to random networks and to the linear the linear threshold model (see [6] for a rigorous proof). Let us assume that the graph is defined via its degree sequence which is i.i.d. distributed according to . Such graphs can be generated using the configuration model [10]. Let , where is the number of edges in the graph. The underlying tree rooted at node can be described by a branching process with the offspring distribution of the root given by . Besides the subsequent generations have offspring distribution

If the degree sequence is such that is finite then, by the strong law of large numbers

Let be the (asymptotic) probability that an edge points to a node with degree . Then for any fixed , the neighbourhood of radius about node , converges in distribution as tends to infinity to a depth- a Galton-Watson tree with a root which has offspring distribution and all other nodes have offspring distribution given by for all . Thus the associated fixed point equation is:

(13)

and we have given by

(14)

As a matter of example, for Erdős-Rényi graphs, the fixed point equation associated with our model is given by (13) and (14) with .

We consider the general threshold model [5]. We have a non-negative random weight on each edge, indicating the influence that exerts on . We consider the symmetric case where and we assume are i.i.d with distribution function . Each node has an arbitrary function defined on subsets of its neighbours set : for any set of neighbours , there is a value between and which is monotone in the sense that if , then . This node chooses a threshold at random from and at time step it becomes active, it plays B, if its set of currently active neighbours satisfies .

5 Conclusion and future work

In this paper we presented two models of marketing wherein individuals, represented by a graph structure, receive payoffs to entice them to adopt a strategy that is different from their initial choice. To this end we initially force a small proportion of nodes to opt for the new strategy and then use an economic model that accounts for the cascading dynamic of adoption. We analyze the evolution of the proportion of agents that switch to the new strategy over time. First, the implications of our results concern marketing strategies in online social networks. More precisely, let be the proportion of forced agents and let the price incurred to force the initial agents. Typically if there is a fixed cost per agent, say , we could take . Let be the proportion of agents choosing at time : . We have which corresponds to the proportion of agents choosing without being initially forced. We denote by the price incurred by the rebates until time . We typically take . Let be the proportion of edges at time t. We denote by the price incurred by the marketing of edges until time . We typically take . Hence the total price of the marketing strategy at time is given by . One can compute the quantities and in function of and . This opens the possibilities of doing an optimal control of the marketing policy.

Finally we remark that the marketing problem that we considered in this paper is just one application of our method. Our approach can indeed be adapted to the analysis of the dissemination of new versions of existing protocols, voting protocols through simple majority rules, i.e., and distributed digital preservation systems [8].

References

  • [1] J. Balogh and B. G. Pittel. Bootstrap percolation on the random regular graph. Random Structures Algorithms, 30(1-2):257–286, 2007.
  • [2] P. Domingos and M. Richardson. Mining the network value of customers. In KDD ’01: Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, pages 57–66, New York, NY, USA, 2001. ACM.
  • [3] A. Ganesh, L. Massoulié, and D. Towsley. The effect of network topology on the spread of epidemics. In Proceedings IEEE Infocom, 2005.
  • [4] S. Janson, T. Łuczak, and A. Rucinski. Random graphs. Wiley-Interscience Series in Discrete Mathematics and Optimization. Wiley-Interscience, New York, 2000.
  • [5] J. Kleinberg. Cascading behavior in networks: Algorithmic and economic issues. In Algorithmic Game Theory (N. Nisan, T. Roughgarden, E. Tardos, V. Vazirani). Cambridge University Press, 2007.
  • [6] M. Lelarge. Diffusion and cascading behavior in random networks. in preparation.
  • [7] V. Mahajan, E. Muller, and F. M. Bass. New product diffusion models in marketing: A review and directions for research. Journal of Marketing, 54(1):1–26, 1990.
  • [8] P. Maniatis, M. Roussopoulos, T. Giuli, D. S. H. Rosenthal, and M. Baker. The lockss peer-to-peer digital preservation system. ACM Transactions on Computer Systems, 23, 2005.
  • [9] S. Morris. Contagion. Rev. Econom. Stud., 67(1):57–78, 2000.
  • [10] R. van der Hofstad. Random graphs and complex networks. In http://www.win.tue.nl/ rhofstad/NotesRGCN2008.pdf, 2008.
  • [11] D. Watts. A simple model of global cascades on random networks. Proceedings of the National Academy of Science, apr 2002.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
139205
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description