TedGram: Twitter Event Detection using Graphbased Methods

Abstract

The proliferation of social networks has made researchers turn to the analysis of these networks. Event detection is one of the important topics in the analysis of social networks, especially Twitter. In this paper, we propose an online graph-based approach, called TedGram, for event detection in Twitter using word embedding techniques and graph partitioning algorithms. In the TedGram model, for each incoming tweet, candidate tweets are gathered from preceding tweets using co-occurrence in entities keywords, and correspondingly the similarity between tweets are computed using the Word Mover’s Distance (WMD) algorithm and pretrained word2vec model. In this regard, the TTI (Tweet Tweet Interaction) graph is computed and updated using an online greedy community detection method based on the Barabási-Albert generative model. Furthermore, we utilize Latent Dirichlet Allocation (LDA) and WMD to combine duplicate communities for detecting and merging duplicate events. Our proposed method is applied to a sample of the Event2012 dataset and is evaluated regarding Precision, Recall, and Fscore. The experimental results show that TedGram performs well against the existing methods.