Scalable Multimodal Factorization for Learning from Very Big Data

Liu, W; Quan, D

Scalable Multimodal Factorization for Learning from Very Big Data

Liu, W

Quan, D

Permalink

Publisher:: Spinger
Publication Type:: Chapter
Citation:: Multimodal Analytics for Next-Generation Big Data Technologies and Applications, 2018
Issue Date:: 2018-12-06

Closed Access

	Filename	Description	Size
	Chapter 10.pdf	Published version	827.17 kB	Adobe PDF	View/Open

Copyright Clearance Process

Recently Added
In Progress
Closed Access

This item is closed access and not available.

Full metadata record

Field	Value	Language
dc.contributor.author	Liu, W https://orcid.org/0000-0002-3003-1313	en_US
dc.contributor.author	Quan, D	en_US
dc.contributor.editor	Seng, KP	en_US
dc.contributor.editor	Ang, L-M	en_US
dc.contributor.editor	Gao, J	en_US
dc.contributor.editor	Liew, AW-C	en_US
dc.date.issued	2018-12-06	en_US
dc.identifier.citation	Multimodal Analytics for Next-Generation Big Data Technologies and Applications, 2018	en_US
dc.identifier.isbn	978-3-319-97598-6	en_US
dc.identifier.uri	http://hdl.handle.net/10453/132179
dc.description.abstract	Recent technology advances in data acquisition bring to re- search communities new opportunities as well as new challenges. They enable researchers to acquire multiple modes of information about the real world. This multimodal data can be naturally and e ciently represented by a multi-way structure, so-called tensors, which can be analyzed to ex- tract the underlying core patterns of the observed data. Multiple datasets obtained from di erent acquisition methods and sensors are increasingly available. The increasing availability of multiple modalities, captured in correlated tensors, provides a complete picture of the whole data patterns. Given large-scale datasets, existing distributed methods for joint analy- sis of multi-dimensional data generated from multiple sources decompose them on several computing nodes following Map-Reduce paradigm. How to improve the performance of Map-Reduce based factorization algorithms as observed data gets bigger is still an open problem. This requires an even more e cient solution that not only reduces communication overhead but also optimizes factors faster. In this book chapter, we provide readers knowledge about Tensor Factor- ization and joint analysis of several correlated Tensors. We propose a Scalable Multimodal Factorization (SMF) algorithm for analyzing corre- lated big multimodal data. It has two key features to enable big multimodal data analysis. Firstly, SMF's design, based on Apache Spark, enables it to have the smallest communication cost. Secondly, its optimized solver converges faster. These key advantages reduce factorization's time com- plexity. As a result, SMF's performance is extremely e cient as the data increases. Con rmed by our experiments with 1 billion known entries, SMF outperforms the currently fastest Coupled Tensor Factorization and Tensor Factorization by 17.8 and 3.8 times, respectively. Compellingly, SMF achieves this speed with the highest accuracy.	en_US
dc.publisher	Spinger	en_US
dc.relation.ispartof	Multimodal Analytics for Next-Generation Big Data Technologies and Applications	en_US
dc.relation.isbasedon	10.1007/978-3-319-97598-6	en_US
dc.title	Scalable Multimodal Factorization for Learning from Very Big Data	en_US
dc.type	Chapter
utslib.location	Switzerland	en_US
pubs.embargo.period	Not known	en_US
pubs.organisational-group	/University of Technology Sydney
pubs.organisational-group	/University of Technology Sydney/Faculty of Engineering and Information Technology
pubs.organisational-group	/University of Technology Sydney/Faculty of Engineering and Information Technology/School of Computer Science
pubs.organisational-group	/University of Technology Sydney/Strength - AAI - Advanced Analytics Institute Research Centre
utslib.copyright.status	closed_access
pubs.consider-herdc	false	en_US
pubs.place-of-publication	Switzerland	en_US
pubs.publication-status	Accepted	en_US

Abstract:

Recent technology advances in data acquisition bring to re- search communities new opportunities as well as new challenges. They enable researchers to acquire multiple modes of information about the real world. This multimodal data can be naturally and e ciently represented by a multi-way structure, so-called tensors, which can be analyzed to ex- tract the underlying core patterns of the observed data. Multiple datasets obtained from di erent acquisition methods and sensors are increasingly available. The increasing availability of multiple modalities, captured in correlated tensors, provides a complete picture of the whole data patterns. Given large-scale datasets, existing distributed methods for joint analy- sis of multi-dimensional data generated from multiple sources decompose them on several computing nodes following Map-Reduce paradigm. How to improve the performance of Map-Reduce based factorization algorithms as observed data gets bigger is still an open problem. This requires an even more e cient solution that not only reduces communication overhead but also optimizes factors faster. In this book chapter, we provide readers knowledge about Tensor Factor- ization and joint analysis of several correlated Tensors. We propose a Scalable Multimodal Factorization (SMF) algorithm for analyzing corre- lated big multimodal data. It has two key features to enable big multimodal data analysis. Firstly, SMF's design, based on Apache Spark, enables it to have the smallest communication cost. Secondly, its optimized solver converges faster. These key advantages reduce factorization's time com- plexity. As a result, SMF's performance is extremely e cient as the data increases. Con rmed by our experiments with 1 billion known entries, SMF outperforms the currently fastest Coupled Tensor Factorization and Tensor Factorization by 17.8 and 3.8 times, respectively. Compellingly, SMF achieves this speed with the highest accuracy.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/132179