Concepção de uma solução escalável para maximização de influência ciente de tópicos em redes sociais.

Abstract

The use of social networks has shown great potential for information diffusion and formation of public opinion. One key problem that has attracted researchers' interest is how to find an initial set of users such that, when given an incentive, they might influence a substantial portion of the network to buy a product, adopt an innovation, or spread news. This problem is known as Influence Maximization. Although major improvements have been made since the íirst solution for this problem was developed, most of these efforts have been concerned on how to solve scalability issues and how to learn the solution parameters. As a result, other key aspects have gained minor interest, such as depending on relationship between social influence and users' topics of interest. Recently, this issue has been addressed as a problem known as Topic-based Influence Maximization, referring to finding a small set of users on a social network that have the ability to influence a substantial portion of users on a given topic. The proposed solutions, however, are not suitable for large-scale social networks and must incorporate mechanisms for determining social influence among users for each topic of interest. Consequently, for these approaches, it becomes difficult or even unfeasible to deal quickly and efficiently with constant changes in the structure of social networks. This problem is particularly relevant when the topics of interest of users and the social influence they exert on each other for every topic are considered together. In this work we propose a scalable solution that makes use of data mining based on an information propagation log, in order to directly select the initial set of influential users on a particular topic without needing to incorporate a previous learning stage of social influence with regard to that topic. As an additional benefit, the targeted seed set also offers an approximation guarantee of the optimal solution. Finally, an experimental evaluation is presented based on datasets containing information propagation data from real social networks where evidence has been found that the proposed solution maintains a trade-off between scalability and accuracy.