Indexed by:
Abstract:
Exploring complicated dynamic spatiotemporal correlations has always been a challenging issue in traffic prediction. Besides, methods that make predictions directly from data with missing values, have received much attention due to the inevitable and pervasive nature of data incompleteness in real scenarios. In this paper, an end-to-end representation learning framework, named spatial- temporal periodical adaptive graph contrastive learning (ST-A-PGCL), is proposed to address such issues. ST-A-PGCL mainly consists of three independent branches to respectively model three long-term periodicities of traffic flow (recent, daily, and weekly periodicities). In each branch, the spatial and tem-poral correlations are extracted by improved adaptive graph convolution network (ImpAdapGCN) and fused seasonal-trend temporal convolution network (FST-TCN) in an encoder, respectively, to obtain hidden representation. Besides, each branch accepts one periodic segment which will be synthetically augmented with different missing patterns to simulate real scenarios (weak communication signal, detector malfunction, area-wide power failure, etc) and generate different views. These views will be fed into a periodical graph contrastive learning (PGCL) module to learn periodical similarity features based on Siamese network to defeat data incompleteness. Bidirectional gate recurrent unit (Bi-GRU) is selected to decode the hidden representations and generate final prediction results. Specifically, the overall framework is trained in end-to-end dual-task (traffic prediction and contrastive learning) process without requiring identifying the position of missing values. Our framework is evaluated across four real-world datasets and twenty baseline models. Experimental results show that the proposed ST-A-PGCL achieves superior prediction performance, especially in long-term prediction tasks with high missing rates.(c) 2023 Elsevier B.V. All rights reserved.
Keyword:
Reprint Author's Address:
Email:
Source :
KNOWLEDGE-BASED SYSTEMS
ISSN: 0950-7051
Year: 2023
Volume: 272
8 . 8 0 0
JCR@2022
ESI Discipline: COMPUTER SCIENCE;
ESI HC Threshold:19
Cited Count:
SCOPUS Cited Count: 19
ESI Highly Cited Papers on the List: 0 Unfold All
WanFang Cited Count:
Chinese Cited Count:
30 Days PV: 0
Affiliated Colleges: