Noise adaptive tensor train decomposition for low-rank embedding of noisy data

Xinsheng Li, K. Selçuk Candan, Maria Luisa Sapino

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Tensor decomposition is a multi-modal dimensionality reduction technique to support similarity search and retrieval. Yet, the decomposition process itself is expensive and subject to dimensionality curse. Tensor train decomposition is designed to avoid the explosion of intermediary data, which plagues other tensor decomposition techniques. However, many tensor decomposition schemes, including tensor train decomposition is sensitive to noise in the input data streams. While recent research has shown that it is possible to improve the resilience of the tensor decomposition process to noise and other forms of imperfections in the data by relying on probabilistic techniques, these techniques have a major deficiency: they treat the entire tensor uniformly, ignoring potential non-uniformities in the noise distribution. In this paper, we note that noise is rarely uniformly distributed in the data and propose a Noise-Profile Adaptive Tensor Train Decompositionmethod, which aims to tackle this challenge. $$\mathtt{NTTD}$$ leverages a model-based noise adaptive tensor train decomposition strategy: any rough priori knowledge about the noise profiles of the tensor enable us to develop a sample assignment strategy that best suits the noise distribution of the given tensor.

Original languageEnglish (US)
Title of host publicationSimilarity Search and Applications - 13th International Conference, SISAP 2020, Proceedings
EditorsShin’ichi Satoh, Lucia Vadicamo, Fabio Carrara, Arthur Zimek, Ilaria Bartolini, Martin Aumüller, Bjorn Por Jonsson, Rasmus Pagh
PublisherSpringer Science and Business Media Deutschland GmbH
Pages203-217
Number of pages15
ISBN (Print)9783030609351
DOIs
StatePublished - 2020
Event13th International Conference on Similarity Search and Applications, SISAP 2020 - Copenhagen, Denmark
Duration: Sep 30 2020Oct 2 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12440 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference13th International Conference on Similarity Search and Applications, SISAP 2020
CountryDenmark
CityCopenhagen
Period9/30/2010/2/20

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of 'Noise adaptive tensor train decomposition for low-rank embedding of noisy data'. Together they form a unique fingerprint.

Cite this