An SVD-based Multimodal Clustering method for Social Event Detection

With the rapid development of social media sites such as Flickr, user-generated multimedia content on the Web has shown an explosive growth in recent years. Social event detection from these large multimedia collections has become one of the hottest topics in analysis of Web content. In this paper, an SVD-based Multimodal Clustering (SVDMC) algorithm is proposed to detect social events from multimodal data. SVDMC is a completely unsupervised approach aiming to take full advantage of the data at hand. Through using the binary adjacency matrix and Singular Value Decomposition (SVD), SVDMC is robust to data incompleteness for datasets in real world. Experiments conducted on the MediaEval Social Event Detection (SED) 2012 dataset demonstrate the effectiveness of the proposed method as well as discriminative power of different features.