Data-driven Outbreak Detection in Social Networks

In social networks, influence such as information, virus, innovations spreads from node to node through the edges of the networks. Understanding this diffusion process is the key to many important real-world applications, for example marketing strategies, virus and pollution control. The problem of outbreak detection and influence maximization ask about selecting a set of “important” nodes (sensor locations, influential bloggers) in a network, from which influence may outbreak or we can detect influence outbursts. These two problems can be modeled as selecting nodes in order to maximize some measures defined over all the sets of nodes. These measures are usually chosen in terms of the number of influenced nodes, the time passed from outbreak until detection and etc. Most traditional studies [6][8] focus on how to maximize the defined measures effectively and efficiently. Research efforts have been made to cast the original problem into the submodular function maximization where greedy hill climbing algorithms can deliver 63% approximation of the optimal. However, what the previous research failed to capture is that in realworld scenarios multiple influences spread through the network simultaneously, and these spread of influences tend to correlate with content of information being spread. Furthermore, nodes may have specific preferences towards specific contents, so that nodes that are chosen to be the “outbreakers” for one type of information cascades may not be the “outbreakers” for information cascades of another content type. This project addresses how to incorporate content into outbreak detection, which we refer to “data-driven outbreak detection”. We propose an adapted model that allows for modeling outbreak detection with content information, and evaluate our model on real world collaboration network and Twitter network. Moreover, we will use the proposed model and machine learning techniques to train classifiers that predict whether a specific piece of content information will cause an outbreak in the network. Finally, we discuss the pros and cons of incorporating content information into the analysis of information outbreak.