Static Seeding and Clustering of LSTM Embeddings to Learn From Loosely Time-Decoupled Events

Humans learn from the occurrence of events at different places and times to predict similar trajectories of events. We define loosely decoupled time (LDT) phenomena as two or more events that could occur in different places and across different timelines but share similarities in the nature of the e...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access Vol. 11; pp. 64219 - 64227
Main Authors: Manasseh, Christian G., Veliche, Razvan, Bennett, Jared, Clouse, Hamilton Scott
Format: Journal Article
Language:English
Published: Piscataway IEEE 2023
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects:
Algorithm design and analysis
artificial intelligence
Clustering
COVID-19
Data models
Diseases
Hidden Markov models
Multivariate analysis
numerical algorithms and problems
Predictive models
Recurrent neural networks
statistical methods
Time series
Time series analysis
Training
ISSN:2169-3536, 2169-3536
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Humans learn from the occurrence of events at different places and times to predict similar trajectories of events. We define loosely decoupled time (LDT) phenomena as two or more events that could occur in different places and across different timelines but share similarities in the nature of the event and the properties of the location. In this work, we improve the use of recurrent neural networks (RNN), particularly long short-term memory (LSTM) networks, to enable AI solutions that generate better time series predictions for LDT. We used similarity measures between the time series based on the time series properties detected by the LSTM and introduced embeddings representing these properties. The embeddings represent the properties of the event, which, coupled with the LSTM structure, can be clustered to identify similar temporally unaligned events. In this study, we explore methods of seeding a multivariate LSTM from time-invariant data related to the geophysical and demographic phenomena modeled by the LSTM. We applied these methods to time-series data derived from COVID-19 detected infection and death cases. We use publicly available socioeconomic data to seed the LSTM models, creating embeddings, to determine whether such seeding improves case predictions. The embeddings produced by these LSTMs are clustered to identify the best-matching candidates for forecasting evolving time series. Applying this method, we showed an improvement in the 10-day moving average predictions of disease propagation at the US County level.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2023.3288487