Temporal Attention Evolutional Graph Convolutional Network for Multivariate Time Series Forecasting

Xinlong Zhao, Liying Zhang, Tianbo Zou, Yan Zhang · 2025

Read Temporal Attention Evolutional Graph Convolutional Network for Multivariate Time Series Forecasting by Xinlong Zhao, Liying Zhang, Tianbo Zou, Yan Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00302.txt.

Introduction

Xinlong Zhao and co-authors present a new method for forecasting many linked time series at once, published in 2025. Their model, called TAEGCN, treats each data node as part of a graph. Unlike older methods, it does not assume that the graph stays fixed over time. It also notes that links can look quite different at fast and slow time scales.

Inside, the model learns time patterns with causal filters and a form of self-attention, then builds a changing graph from those patterns. Tests use two public road traffic data sets, METR-LA and PEMS-BAY. The authors report that their method beats prior methods on these tests, which may help with traffic planning and other forecasts.

Catalog record

CREATORXinlong Zhao, Liying Zhang, Tianbo Zou, Yan Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-00302

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Temporal Attention Evolutional Graph Convolutional Network for Multivariate Time Series Forecasting Xinlong Zhao, Liying Zhang, Tianbo Zou, Yan Zhang Abstract Multivariate time series forecasting enables the prediction of future states by leveraging historical data, thereby facilitating decision-making processes. Each data node in a multivariate time series encompasses a sequence of multiple dimensions. These nodes exhibit interdependent relationships, forming a graph structure. While existing prediction methods often assume a fixed graph structure, many real-world scenarios involve dynamic graph structures. Moreover, interactions among time series observed at different time scales vary significantly. To enhance prediction accuracy by capturing precise temporal and spatial features, this paper introduces the Temporal Attention Evolutional Graph Convolutional Network (TAEGCN). This novel method not only integrates causal temporal convolution and a multi-head self-attention mechanism to learn temporal features of nodes, but also construct the dynamic graph structure based on these temporal features to keep the consistency of the changing in spatial feature with temporal series. TAEGCN adeptly captures temporal causal relationships and hidden spatial dependencies within the data. Furthermore, TAEGCN incorporates a unified neural network that seamlessly integrates these components to generate final predictions. Experimental results conducted on two public transportation network datasets, METR-LA and PEMS-BAY, demonstrate the superior performance of the proposed model. keywords Multivariate Time Series Forecasting, Graph Convolutional Network, Multi-head Self Attention, Dynamic Mapping 1 Introduction Multivariate Time Series (MTS) encapsulates a continuous time span, reflecting changes in multiple variables over this duration, such as air quality, commodity prices, among others. MTS embodies a collection of data with inherent dependencies, often represented as a graph structure where nodes denote variables. Classical forecasting methods for MTS encompass statistical models and machine learning techniques. Statistical models, exemplified by the Autoregressive Integrated Moving Average (ARIMA) model, offer high computational efficiency and interpretability across various domains Guo et al. 2019 ; Liu and Liu 2009 . However, the emergence of deep learning models has garnered significant interest due to their adeptness in nonlinear modeling and resilience to diverse data distributions Lee et al. 2021 . Consequently, researchers increasingly leverage deep learning approaches for MTS forecasting. To exploit MTS characteristics fully, some researchers employ Convolutional Neural Networks (CNNs) to capture temporal proximity information, while Recurrent Neural Networks (RNNs) are applied along the temporal axis. For instance, Wu et al. proposed a Convolutional Long Short-Term Fusion Prediction (CLTFP) architecture, combining Long Short-Term Memory (LSTM) and 1-dimensional convolution to predict short-term traffic conditions Hochreiter and Schmidhuber 1997 ; Wu and Tan 2016 . Although CLTFP adopts a straightforward approach, it pioneers the alignment of temporal and spatial regularities. However, the rigid structure of regular convolutions confines the model’s applicability to grid-based data structures like images or videos, rather than accommodating more diverse structures like graphs. Additionally, RNNs entail iterative training for sequence learning, leading to error accumulation and computational burden. LSTNet Lai et al. 2018 and TPA-LSTM Shih et al. 2019 stand as classical models in MTS prediction, blending convolutional and recurrent neural networks to capture intra- and inter-series correlations. Nonetheless, the non-Euclidean spatial relationships among nodes pose challenges for CNNs’ global aggregation in accurately capturing variable correlations. To address this challenge, recent exploratory research has delved into leveraging Graph Convolutional Networks (GCNs), which are adept at processing non-Euclidean spaces, to tackle issues encountered in multivariate time series (MTS) prediction where Convolutional Neural Networks (CNNs) falter. GCNs have witnessed widespread application in MTS prediction across various domains. They construct a graph structure wherein each MTS constitutes a node, and edges accurately represent interconnections between different nodes, thereby forming a graph structure. Literature Shang et al. 2021 simplifies the dual-layer program issue and samples discrete graph structures from Bernoulli distributions. Designing suitable graph structures to model correlations between MTS elements and time steps has emerged as a pivotal research focus in this domain. Initially, Li et al. 2017 integrated GCNs with gated recursive units to make predictions and introduced a manually crafted adjacency matrix to depict correlations based on node distances. Subsequently, Wu et al. 2019 contended that predefined graph structures fail to reflect genuine connections and proposed the use of a self-learning adjacency matrix during training. The single-layer GCN serves as a first-order approximation of Chebynet Defferrard et al. 2016 , realized by stacking multiple layers to approximate a high-order polynomial filter. Models for MTS prediction, which comprehend both spatial node characteristics and temporal node evolution, necessitate amalgamating temporal and spatial data information. A model that can concurrently capture temporal and spatial correlations within historical data is termed a Spatio-Temporal Model (STM). Literature Wu et al. 2019 ; Wu et al. 2020 ; Gu et al. 2022 utilized GCNs for spatial information extraction and Temporal Convolutional Network (TCN) for temporal information extraction. Another study Bai et al. 2020 employed polynomial graph convolution filters and RNNs for extracting temporal dimension information. Additionally, Zheng et al. 2020 proposed a graph attention module for spatial information transfer and combined temporal attention or multi-graph parallel modeling to jointly learn spatio-temporal representations. Spatio-temporal separation models utilize distinct models for extracting temporal and spatial information, whereas spatio-temporal joint models treat time series as directed line graphs that can integrate with graph structures, thereby reducing modeling degrees of freedom Pan et al. 2020 . Temporal feature extraction in literature Liang et al. 2023 combines multinomial self-attention with long- and short-term time series analyses. In scenarios lacking an existing adjacency matrix for a given graph, the graph structure must be constructed initially. However, solely constructing the graph structure based on a particular metric is deemed inadequate Guo et al. 2020 . Hence, Graph Structure Learning (GSL) assumes a pivotal role in modeling complex networks or graph neural networks. Traditional GSL methods rely on statistical or optimization principles Roddenberry et al. 2021 , encompassing metric-based methods, probabilistic methods, and direct optimization methods Zhu et al. 2021 . Metric-based methods Wu et al. 2020 ; Bai et al. 2020 necessitate initializing an embedding vector for each node, independent of the node’s features. During training, the model optimizes this vector and constructs the graph’s adjacency matrix using a metric function. While requiring fewer parameters, these methods suffer from slow convergence. On the other hand, the probability-based GSL method Elinas et al. 2020 learns conditional probabilities between node pairs of features, combining them with prior graphs to construct the adjacency matrix, albeit without computational or parameter advantages. Additionally, Ye et al. 2022 proposed learning graph structures based on temporal sequences. The exploration of using machine learning to jointly infer graph structures and train predictive models in an end-to-end manner is a current focal point. The complexity inherent in multivariate time series data poses significant challenges to improving prediction accuracy, driving the ongoing development of new prediction methods as a primary research direction. While recent advancements in graph convolutional networks (GCNs) applied to multivariate time series prediction have yielded notable successes through long- and short-term time series analysis and inter-node spatial feature extraction, several critical challenges persist, necessitating further research: 1. Inefficient Time Dimension Feature Extraction: Efficiently extracting key temporal information from each element within a multivariate time series is crucial. Since individual element sequence data may not adequately characterize the overall dataset, appropriate extraction of temporal information from multivariate time series is imperative. However,

Read the complete text