Tensor Methods for Data Modeling

Association for Women in Mathematics (AWM) Annual Symposium, 2023


The Association for Women in Mathematics (AWM) 2023 Research Symposium featured a Session on Tensor Methods for Data Modeling organized by Anna Konstorum (Center for Computing Sciences, Institute for Defense Analyses). The goal of the session was to bring together speakers working in on a variety of tensor analysis methods with the aim to improve their applicability to real-world problems. Slides shared here have been made available by the speakers.


Abstract


Research into applications of tensor decomposition methods to real-world data has been accelerating, driven by the increasing complexity of data sets arising from a broad range of fields, including the biomedical, network, and physical and social sciences. Tensor methods can aid in low-dimensional sample and feature embedding and clustering, as well as data compression. Current challenges in deployment of tensor decompositions for data analytics include questions of existence and uniqueness of optimal solutions, rank and model selection, computational complexity, and data imputation. The goal of this session is to bring together speakers that develop and implement tensor methods on real-world datasets, including aspects relating to the theoretical and/or computational challenges, to foster collaboration and discussion.

Talks


Hierarchical nonnegative tensor factorizations and applications

Speaker: Jamie Haddock (Harvey Mudd College)
Co-authors: Deanna Needell, Joshua Vendrow

Slides

Abstract. Nonnegative matrix factorization (NMF) has found many applications including topic modeling and document analysis. Hierarchical NMF (HNMF) variants are able to learn topics at various levels of granularity and illustrate their hierarchical relationship. Recently, nonnegative tensor factorization (NTF) methods have been applied in a similar fashion in order to handle data sets with complex, multi-modal structure. Hierarchical NTF (HNTF) methods have been proposed, however these methods do not naturally generalize their matrix-based counterparts. Here, we propose a new HNTF model which directly generalizes a HNMF model special case, and provide a supervised extension. Our experimental results show that this model more naturally illuminates the topic hierarchy than previous HNMF and HNTF methods.

Recovering Noisy Tensors from Double Sketches

Speaker: Anna Ma (UC Irvine)
Co-authors: Yizhe Zhu, Dominik Stoeger

Slides

Abstract. Machine learning tasks often utilize vast amounts of data where data can be represented in various ways. Instead of representing data as one or two-dimensional arrays, one can instead use multi-dimensional arrays to describe more complex relationships between elements in a data set. Moving beyond the matrix or vector representation of data requires using tensors and tensor operations. In this talk, we focus on using the tensor-tensor t-product for recovering tensors of total and assumably low tensor rank. In particular, we will explore using random Gaussian sketches for recovering tensors and present provable guarantees for recovery.

Robust Tensor CUR: Rapid Low-Tucker-Rank Tensor Recovery with Sparse Corruptions

Speaker: Longxiu Huang (Michigan State University)
Co-authors: HanQin Cai, Zehan Chao, Longxiu Huang, and Deanna Needell

Slides

Abstract. We study the problem of tensor robust principal component analysis (TRPCA), which aims to separate an underlying low-multilinear-rank tensor and a sparse outlier tensor from their sum. This work proposes fast algorithms, called Robust Tensor CUR(RTCUR), for large-scale non-convex TRPCA problems under the Tucker rank setting. RTCUR is developed within a framework of alternating projections that projects between the set of low-rank tensors and the set of sparse tensors. We utilize the recently developed tensor CUR decomposition to substantially reduce the computational complexity in each projection. In addition, we develop four variants of RTCUR for different application settings. We demonstrate the effectiveness and computational advantages of RTCUR against state-of-the-art methods on both synthetic and real-world datasets.

Tensor BM-Decomposition for Compression and Analysis of Spatio-Temporal Third-order Data

Speaker: Fan Tian (Tufts University)
Co-authors: Misha E. Kilmer, Eric Miller, Abani Patra

Slides

Abstract. We introduce a third-order tensor decomposition framework based on a ternary multiplication named Bhattacharya-Mesner (BM) product and its corresponding notion of rank. We describe an iterative algorithm for computing a low BM-rank approximation to a given third-order data array. Moreover, we will demonstrate our decomposition framework for video processing applications. Our method appears to improve on other matrix and tensor-based methods with smaller approximation error for the same level of compression.