Mesoscale Anisotropically-Connected Learning

Qi Tan, Yang Liu, Jiming Liu*

*Corresponding author for this work

Research output: Chapter in book/report/conference proceedingConference proceedingpeer-review

Abstract

Predictive spatio-temporal analytics aims to analyze and model the data with both spatial and temporal attributes for future forecasting. Among various models proposed for predictive spatio-temporal analytics, the recurrent neural network (RNN) has been widely adopted. However, the training of RNN models becomes slow when the number of spatial locations is large. Moreover, the structure of RNN is unable to dynamically adapt to incorporate new covariates or to predict the target variables with varying spatial dimensions. In this paper, we propose a novel method, named Mesoscale Anisotropically-Connected Learning (MACL), to address the aforementioned limitations in RNN. For efficient training, we group the dataset into clusters (which refers to the mesoscale) along the spatial dimension according to the spatial adjacency and develop individual prediction module for each cluster. Then we design an anisotropic information exchange mechanism (i.e., the information exchange is not symmetric), to allow the prediction modules leveraging state information from nearby clusters for enhancing the prediction accuracy. Furthermore, for timely adaptation, we develop a local updating strategy for adapting the learning model to incorporate new covariates and the target variables with varying spatial dimensions. Experimental results on a real-world prediction task demonstrate that our method can be trained faster and more accurate than existing methods. Moreover, our method is flexible to incorporate new covariates and target variables of varying spatial dimensions, without sacrificing the prediction accuracy.

Original languageEnglish
Title of host publicationFoundations of Intelligent Systems
Subtitle of host publication25th International Symposium, ISMIS 2020, Graz, Austria, September 23–25, 2020, Proceedings
EditorsDenis Helic, Martin Stettinger, Alexander Felfernig, Gerhard Leitner, Zbigniew W. Ras
Place of PublicationCham
PublisherSpringer
Pages171-180
Number of pages10
Edition1st
ISBN (Electronic)9783030594916
ISBN (Print)9783030594909
DOIs
Publication statusPublished - 17 Sept 2020
Event25th International Symposium on Methodologies for Intelligent Systems, ISMIS 2020 - Graz, Austria
Duration: 23 Sept 202025 Sept 2020

Publication series

NameLecture Notes in Computer Science
Volume12117
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349
NameISMIS: International Symposium on Methodologies for Intelligent Systems

Conference

Conference25th International Symposium on Methodologies for Intelligent Systems, ISMIS 2020
Country/TerritoryAustria
CityGraz
Period23/09/2025/09/20

Scopus Subject Areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint

Dive into the research topics of 'Mesoscale Anisotropically-Connected Learning'. Together they form a unique fingerprint.

Cite this