2:20 PM - 3:05 PM
Chair:
Matthias Bolten, Universität Wuppertal, Germany
Constrained Low Rank Approximation (CLRA) is a powerful foundation for important data analytic tasks such as topic modeling and community detection. Some advantages of CLRA include scalable algorithms and software based on advances in numerical linear algebra and parallel computing. Nonnegativity constraints allow more judicious formulation, interpretable results, and effective methods whether the input data is in feature-data relationship as in the Nonnegative Matrix Factorization (NMF) or data-data relationship as in spectral clustering or symmetric NMF (SymNMF). A common foundation of CLRA also allows a hybrid method for information fusion called JointNMF from merging the objective functions of the NMF and SymNMF for multi-view data sets with both content and connection information.
In this talk, a distributed memory parallel algorithm and software framework, PLANC (Parallel Low-rank Approximation with Nonnegativity Constraints), is described for nonnegativity constrained matrix and tensor low rank approximations. PLANC uses parallel distributions and algorithms designed to optimize communication and computation, allows extension regarding data characteristics, algorithm, constraints, and architecture beyond the above mentioned NMF variants. Efficiency and scalability of PLANC and some new knowledge discoveries from large real-world data sets are presented.
Haesun Park
| PP22 Home 2022 | Program | Speaker Index |