A Latent Space Approach to Dynamic Embedding of Cooccurrence Data.pdfVIP

A Latent Space Approach to Dynamic Embedding of Co-occurrence Data Purnamrita Sarkar Sajid M. Siddiqi Geoffrey J. Gordon Machine Learning Department Robotics Institute Machine Learning Department Carnegie Mellon University Carnegie Mellon University Carnegie Mellon University Pittsburgh, PA 15213 Pittsburgh, PA 15213 Pittsburgh, PA 15213 Abstract tions of document topics, co-authorship patterns, or recommender systems; the resulting embeddings are We consider dynamic co-occurrence data, useful for tasks like clustering, visualization, informa- such as author-word links in papers pub- tion retrieval, and exploratory data analysis. lished in successive years of the same con- Well-known algorithms for the embedding problem in- ference. For static co-occurrence data, re- clude MDS (Borg Groenen, 1997), IsoMap (Tenen- searchers often seek an embedding of the baum et al., 2000), Locally Linear Embedding entities (authors and words) into a low- (LLE) (Roweis Saul, 2000) and spectral cluster- dimensional Euclidean space. We general- ing (Ng et al., 2001). (Raftery et al., 2002) intro- ize a recent static co-occurrence model, the duced a model similar to MDS in which entities are CODE model of Globerson et al. (2004), to associated with locations in p-dimensional space, and the dynamic setting: we seek coordinates fo