Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks
By Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, Chengqi Zhang. Knowledge Discovery and Data Mining, 2020
2,163 citations recorded by the source corpus.
Modeling multivariate time series has long been a subject that has attracted researchers from a diverse range of fields including economics, finance, and traffic. A basic assumption behind multivariate time series forecasting is that its variables depend on one another but, upon looking closely, it is fair to say that existing methods fail to fully exploit latent spatial dependencies between pairs of variables. In recent years, meanwhile, graph neural networks (GNNs) have shown high capability in handling relational dependencies. GNNs require well-defined graph structures for information propagation which means they cannot be applied directly for multivariate time series where the dependencies are not known in advance. In this paper, we propose a general graph neural network framework designed specifically for multivariate time series data. Our approach automatically extracts the uni-directed relations among variables through a graph learning module, into which external knowledge like variable attributes can be easily integrated. A novel mix-hop propagation layer and a dilated inception layer are further proposed to capture the spatial and temporal dependencies within the time series. The graph learning, graph convolution, and temporal convolution modules are jointly learned in an end-to-end framework. Experimental results show that our proposed model outperforms the state-of-the-art baseline methods on 3 of 4 benchmark datasets and achieves on-par performance with other approaches on two traffic datasets which provide extra structural information.
Subfields: Building and Construction, Management Science and Operations Research, Signal Processing
View Paper: https://www.semanticscholar.org/paper/75e924bd79d27a23f3f93d9b1ab62a779505c8d2
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Building and Construction
In Management Science and Operations Research
- Bernie Wang
- Tim Januschowski
- Jan Gasthaus
- Syama Sundar Rangapuram
- Laurent Callot
- Junchi Yan
- Danielle C. Maddix
- Yunhao Zhang
In Signal Processing
- Bernie Wang
- Tim Januschowski
- Jan Gasthaus
- Syama Sundar Rangapuram
- Danielle C. Maddix
- Junchi Yan
- Yunhao Zhang
- Laurent Callot
Citations in the corpus
- FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective (2023, neutral): “Since the prior graph structure is usually unknown in most multivariate time series scenarios [10, 2, 13], and the elements of X t are spatially or temporally correlated with each other because of time lag effect [28], we assume all nodes in the hypervariate graph G t are fully-connected.”
- Deep Learning for Time Series Forecasting: Tutorial and Literature Survey (2020, methodology): “An embedding is learned for each dimension, and similarity scores are computed between every two dimension using these embeddings from which the adjacency matrix is obtained, either by taking the Ktop edges [43, 193] or sampling from them [163, 102].”
- Less Is More: Fast Multivariate Time Series Forecasting with Light Sampling-oriented MLP Structures (2022, methodology): “In this case, we use the interdependency modeled by MTGNN [24] as the “ground truth” to measure the quality of the interdependency modeled by LightTS.”
- Learning Graph Structures With Transformer for Multivariate Time-Series Anomaly Detection in IoT (2021, methodology): “Graph Convolutional Networks (GCNs) [21], [22], [23], [24] have recently revealed discriminative power in learning graph representations due to their permutation-invariance, local connectivity, and compositionality [21], [25].”
- Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting (2023, methodology): “…neural models explicitly capture the cross-dimension dependency, i.e. preserving the information of dimensions in the latent feature space and using convolution neural network (CNN) (Lai et al., 2018) or graph neural network (GNN) (Wu et al., 2020; Cao et al., 2020) to capture their dependency.”
- PDFormer: Propagation Delay-aware Dynamic Long-range Transformer for Traffic Flow Prediction (2023, background): “…were widely used for traffic prediction (Li et al. 2018; Yu, Yin, and Zhu 2018; Wu et al. 2019; Ji et al. 2020; Chen et al. 2020; Ye et al. 2021; Wu et al. 2020; Zhang et al. 2021; Song et al. 2020; Li and Zhu 2021; Oreshkin et al. 2021; Han et al. 2021; Ji et al. 2022; Fang et al. 2021; Choi…”
- Spatio-Temporal Meta-Graph Learning for Traffic Forecasting (2022, neutral): “Models including MTGNN (Wu et al. 2020), AGCRN (Bai et al. 2020), GTS (Shang, Chen, and Bi 2021) fall into this category, integrating MTS and traffic forecasting with Graph Structure Learning (GSL) (Zhu et al. 2021).”
- Dynamic Graph Convolutional Recurrent Network for Traffic Prediction: Benchmark and Solution (2021, methodology): “STGCN [34] Distance-based graph GCN CNN Time ASTGCN [9] Binary graph GCN + Attention CNN + Attention Time STSGCN [23] Binary graph GCN GCN Time LSGCN [12] Binary graph GCN + Attention CNN Time Graph WaveNet [30] Adaptive graph + Distance-based graph GCN CNN Time SLCNN [36] Adaptive graph GCN CNN Tim”
- Graph Neural Network for Traffic Forecasting: A Survey (2021, background): “Dataset Name Relevant Studies METR-LA [3], [4], [22], [25], [26], [27], [35], [38], [43], [44], [55], [57], [71], [72], [80], [105], [110], [117], [125], [129], [132], [137], [138], [141], [142], [143], [144], [145], [146], [148], [149], [153], [155], [159], [160], [172] PeMS all [28], [147] PeMS-BA”
- A Survey on Graph Neural Networks for Time Series: Forecasting, Classification, Imputation, and Anomaly Detection (2023, methodology): “Other representative examples include MTGNN [51] and MTGODE [23].”
- Spatio-Temporal Graph Neural Networks for Predictive Learning in Urban Computing: A Survey (2023, methodology): “For example, MTGNN [171] employs multiple TCN with various kernel sizes for learning temporal dependencies in different scales, DMVST-VGNN [19] utilizes TCN and Transformer jointly for long-short range temporal learning.”
- Discrete Graph Structure Learning for Forecasting Multiple Time Series (2021, background): “MTGNN (Wu et al., 2020) parameterizes the graph as a degree-k graph, which is learned end-to-end with a GNN for forecasting time series.”
- Dynamic and Multi-faceted Spatio-temporal Deep Learning for Traffic Speed Forecasting (2021, methodology): “Except for our model, Graph Wavenet and MTGNN perform well generally which might benefit from that they adopt adaptive graph to model relationships between nodes, indicating that adaptive graph based methods could effectively exploit valuable but latent spatial dependencies from historical traffic d”
- Graph Neural Networks for Intelligent Transportation Systems: A Survey (2023, background): “Wu et al. [37] also tried to consider a dynamic and flexible graph structure instead of assuming a fixed one.”
- Spatial-Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting (2022, methodology): “HI MAE 42.33 42.35 42.37 42.36 49.02 49.03 49.06 49.04 34.55 34.57 34.59 34.57 3.06 3.06 3.05 3.06 92.44 92.58 92.79 92.58 RMSE 61.64 61.66 61.67 61.66 71.16 71.18 71.20 71.18 50.41 50.43 50.44 50.43 7.05 7.05 7.03 7.04 167.00 167.05 167.21 167.07 MAPE 29.90% 29.92% 29.96% 29.92% 22.73% 22.75% 22.79”
- Anomaly Detection on Attributed Networks via Contrastive Self-Supervised Learning (2021, methodology): “Currently, GNNs have been applied to various research fields, such as time-series prediction [44], [45], hyperspectral”
- Unsupervised Deep Anomaly Detection for Multi-Sensor Time-Series Signals (2021, methodology): “Moreover, the multiple time series signal usually has some kinds of co-relations, [53] propose a method to learn the relation graph on multiple time series.”