Structured Sequence Modeling with Graph Convolutional Recurrent Networks

By Youngjoo Seo, Michaël Defferrard, Pierre Vandergheynst, Xavier Bresson. 2016

984 citations recorded by the source corpus.

This paper introduces Graph Convolutional Recurrent Network (GCRN), a deep learning model able to predict structured sequences of data. Precisely, GCRN is a generalization of classical recurrent neural networks (RNN) to data structured by an arbitrary graph. The structured sequences can represent series of frames in videos, spatio-temporal measurements on a network of sensors, or random walks on a vocabulary graph for natural language modeling. The proposed model combines convolutional neural networks (CNN) on graphs to identify spatial structures and RNN to find dynamic patterns. We study two possible architectures of GCRN, and apply the models to two practical problems: predicting moving MNIST data, and modeling natural language with the Penn Treebank dataset. Experiments show that exploiting simultaneously graph spatial and dynamic information about data can improve both precision and learning speed.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Geometric Matrix Completion with Recurrent Multi-Graph Neural Networks (2017, methodology): “This model was later extended to deal with dynamic data (Seo et al., 2016).”
Recurrent Event Network for Reasoning over Temporal Knowledge Graphs (2019, methodology): “EvolveGCN [18] and GCRN [21] introduce a RNN to update node representations across different time stamps.”
Robust Spatial Filtering With Graph Convolutional Neural Networks (2017, methodology): “Graph signals filtered with this technique are fed through an LSTM in [19] to model sequences.”
Dynamic Graph Convolutional Recurrent Network for Traffic Prediction: Benchmark and Solution (2021, methodology): “Following [4, 15, 21], we replace the matrix multiplications in GRU with dynamic graph convolution modules and get DGCRM, denoted as follows:”
Spatio-Temporal Meta-Graph Learning for Traffic Forecasting (2022, neutral): “9) ∼ 12) are built based upon GCRN (Seo et al. 2018; Li et al. 2018).”
DeepSphere: Efficient spherical Convolutional Neural Network with HEALPix sampling for cosmological applications (2018, neutral): “By combining graph convolutional layers and recurrent layers [66], they can, for example, model structured time series such as tra ﬃ c on road networks [67], or recursively complete matrices for recommendation [68].”
Optimized Graph Convolution Recurrent Neural Network for Traffic Prediction (2021, methodology): “The GCGRU network [47] is an improved version of GRU based on GCN and has been used in trafﬁc prediction [46].”
Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks (2020, methodology): “spatial-temporal graph neuralnetworks, spatialdependencies amongnodes arecaptured by graph convolutions, while temporal dependencies among historical states are preserved by recurrent neural networks [13, 17] or 1D convolutions [22, 23]. Although existing spatial-temporal graph neural networks have ”
Learning from History: Modeling Temporal Knowledge Graphs with Sequential Copy-Generation Networks (2020, methodology): “…(Dasgupta, Ray, and Talukdar 2018), TA-DistMult (Garc´ıa-Dur´an, Dumancic, and Niepert 2018), Know-Evolve+MLP (Trivedi et al. 2017), DyRep+MLP (Trivedi et al. 2018), R-GCRN+MLP (Seo et al. 2018), and RE-NET (Jin et al. 2019), whereof RE-NET has offered SOTA performance on all of the benchmarks.”
Relational Representation Learning for Dynamic (Knowledge) Graphs: A Survey (2019, methodology): “Model 1 of [201] uses this approach where M is the GCN proposed in [59] and the RNN is a standard LSTM. Narayan and Roe [166] also use this approach with M being the GCN proposed in [174] and the RNN being a standard LSTM. Manessi et al. [155] modify this approach slightly by (mainly) adding…”
Incremental Training of Graph Neural Networks on Temporal Graphs under Distribution Shift (2020, background): “Regarding dynamic graphs , several works consider a ﬁxed vertex set and stack recurrent modules on top of graph convolutional modules [23,15].”
GaAN: Gated Attention Networks for Learning on Large and Spatiotemporal Graphs (2018, methodology): “(Seo et al., 2016) proposed Graph Convolutional Recurrent Neural Network (GCRNN), which replaced the fully-connected layers in LSTM (Hochreiter and Schmidhuber, 1997) with the ChebNet operator (Defferrard et al., 2016), and applied it to a synthetic video prediction task.”