Estimating Ground Shaking Regions with Social Media Propagation Trees

摘要

The Mercalli scale of quake damages is based on perceived effects and it has a strong dependence on observers. Recently, we pro-posed a method for ground shaking intensity estimation based on lexical features extracted from tweets, showing good performance in terms of mean absolute error (MAE). One of the flaws of that method is the detection of the region of interest, i.e., the area of a country where the quake was felt. Our previous results showed enough recall in terms of municipality recovery but a poor performance in terms of accuracy. One of the reasons that help to explain this effect is the presence of data noise as many people comment or confirm a quake in areas where the event was unperceived. This happens because people get awareness of an event by watching news or by word-of-mouth propagation. To alleviate this problem in our earthquake detection system we study how propagation features behave in a region of interest estimation task. The intuition behind our study is that the patterns that characterize a word-of-mouth propagation differ from the patterns that characterize a perceived event. If this intuition is true, we expect to separate both kinds of propagation modes. We do this by computing a number of features to represent propagation trees. Then, we trained a learning algorithm using our features in the specific task of region of interest estimation. Our results show that propagation features behave well in this task, outperforming lexical features in terms of accuracy.