Organizing multimedia data socially

Social network sites such as Orkut, MySpace, Facebook, Flickr, and YouTube are flourishing. On these sites, users create communities, upload/share user generated content such as blogs, photos and videos, and interact with each other directly or indirectly. For analyzing and organizing multimedia data, social networks provide useful signals additional to traditional perceptual signals such as color, texture, shape, and motion. In this talk, I will first explain what some social signals can be useful. I will then present algorithms that can fuse these social signals with text and perceptual features. To deal with a large amount of data and rapidly growing social networks, we have recently developed a number of parallel, online algorithms. I will present parallel implementation of Support Vector Machines, PF-Growth (for association mining), spectral clustering algorithm (including SVD and k-mean), Latent Dirichlet Allocation (LDA), and combinational collaborative filtering.