Speakers:

Jon Kleinberg
Department of Computer Science
Cornell University
kleinber@cs.cornell.edu
http://www.cs.cornell.edu/home/kleinber/

Andrew Tomkins
Principles and Methodologies Group
IBM Almaden Research Center
tomkins@almaden.ibm.com
http://www.almaden.ibm.com/cs/people/tomkins/

Algorithms for data analysis draw heavily on both discrete and continuous techniques, often in combination. Combinatorial optimization plays a natural role in methods for clustering and classification: there are a variety of ways to describe 'good' clusterings of a dataset as optimizing a discrete objective function, and this leads to heuristics involving local search and linear programming relaxations. In a different vein, methods based on eigenvectors and the singular value decomposition have been employed for both clustering and approximating high-dimensional data; such approaches form the basis of the Latent Semantic Indexing technique in information retrieval and the current generation of link-based ranking algorithms for Web search. Our understanding of the power of all these methodologies has benefited from the study of probabilistic generative models for large datasets and networks; such models provide a setting in which to rigorously analyze data analysis algorithms, and they also can be used to posit 'simple' explanations for phenomena that are observed across a diverse range of datasets.

LIST OF CONFIRMED PARTICIPANTS