Regularized Non-Negative Matrix Factorization for Dynamic and Relational Data

• Activity Number: 465
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Shawn Mankad*+ and George Michailidis
• Companies: and University of Michigan
• Keywords: Space-time Clustering ; Non-negative matrix factorization ; Graph Regularization ; Multifactorial Data Analysis ; Dynamic Network Analysis ; Visual Analytics

Abstract

Data involving repeated measurements of several variables over time may exhibit correlations among variables, samples, and between time points. The discovery of these underlying, meaningful relations is important to a wide variety of areas such as psychology, signal processing, finance, among others. Common methods such as independent component analysis, factor analysis and others provide a great deal of flexibility, but may not utilize properties of the generating process and hence are susceptible to noise. We discuss the development and application of a novel regularization framework for non-negative matrix factorization for feature extraction and latent source separation in dynamic and relational data. Our framework allows for the incorporation of additional knowledge and model information, resulting in decompositions that are interpretable and robust to noise. We'll illustrate the framework on dynamic graph series drawn from citation and world trade data.