Burst Detection in a Sequence of Tweets Based on Information Diffusion Model

We propose a method of detecting the period in which a burst of information diffusion took place from an observed diffusion sequence data over a social network and report the results obtained by applying it to the real Twitter data. We assume a generic information diffusion model in which time delay associated with the diffusion follows the exponential distribution and the burst is directly reflected to the changes in the time delay parameter of the distribution (inverse of the average time delay). The shape of the parameter change is approximated by a series of step functions and the problem of detecting the change points and finding the values of the parameter is formulated as an optimization problem of maximizing the likelihood of generating the observed diffusion sequence. Time complexity of the search is almost proportional to the number of observed data points (possible change points) and very efficient. We apply the method to the real Twitter data of the 2011 To-hoku earthquake and tsunami, and show that the proposed method is by far efficient than a naive method that adopts exhaustive search, and more accurate than a simple greedy method. Two interesting discoveries are that a burst period between two change points detected by the proposed method tends to contain massive homogeneous tweets on a specific topic even if the observed diffusion sequence consists of heterogeneous tweets on various topics, and that assuming the information diffusion path is a line shape tree can give a good approximation of the maximum likelihood estimator when the actual diffusion path is not known.