Joint Embedding of Disparate Information with Applications in Text Analysis

• Activity Number: 442
• Type: Invited
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): David Marchette*+ and Carey Priebe
• Companies: Naval Surface Warfare Center and The Johns Hopkins University
• Keywords: multidimensional scaling ; disparate data fusion ; dimensionality reduction ; text data mining ; implicit translation

Abstract

The analysis of scientific documents often results in the extraction of features of very different type. For example, in addition to standard text analysis features such as used in bag-of-words models, features extracted from natural language processing methods, or similar text-based features, one may extract graphs such as co-authorship or citation networks, features extracted from images or figures within the documents. One way to perform inference on such different and complex data is to embed the data into a single space so that the different types of information can be combined in a useful way. We discuss several approaches to this problem, derived from a dissimilarity framework and utilizing multidimensional scaling ideas to define the embeddings. We will illustrate these techniques on several text document datasets.