Research map / Time series / Long-term forecasting
Motif-aware Graph Neural Networks for Networked Time Series Imputation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Motif-GNN, a graph neural network for networked time series imputation that fuses motif-induced and motif-based adjacency matrices via motif-wise multi-view attention and combines GNN with a gated GRU fusion layer.
Problem
Imputing missing values in networked time series (multivariate sensor series on a graph) when existing GNN methods capture only low-order neighborhood structure and do not identify which higher-order motifs matter.
Approach
Builds motif-induced and motif-based adjacency matrices from small graph motifs, fuses them with learned attention weights over motifs, and integrates the resulting structural embeddings with a GRU in a gated fusion layer for autoregressive imputation.
Key results
- Motif-GNN achieves MAE 15.43 on Air Quality under general missing vs. 19.49 for PoGeVon, roughly 21% lower.
- Overall improvement across four datasets is 16–21% in MAE and 18–19% in MRE relative to the best baseline.
- Ablation shows removing motif-induced matrices, motif-based matrices, or multi-view attention each degrades MAE on all datasets (e.g., Air Quality 15.43 -> 17.49, 18.16, 19.21).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Motif-GNN instead of PoGeVon | Networked time series imputation · Air Quality, general missing | MAE: 19.49 ± 1.10 -> 15.43 ± 0.32 | Helped |
| Use Motif-GNN instead of PoGeVon | Networked time series imputation · PeMS-LA, general missing | MAE: 23.91 ± 0.25 -> 18.21 ± 0.12 | Helped |
| Use Motif-GNN instead of PoGeVon | Networked time series imputation · PeMS-BA, general missing | MAE: 22.19 ± 0.05 -> 17.25 ± 0.16 | Helped |
| Use Motif-GNN instead of PoGeVon | Networked time series imputation · PeMS-SD, general missing | MAE: 18.99 ± 0.11 -> 15.87 ± 0.14 | Helped |
| Add motif-induced matrices (ablation) | Motif-GNN · Air Quality, MCAR | MAE: 17.49 ± 0.32 -> 15.43 ± 0.32 | Helped |
| Add motif-induced matrices (ablation) | Motif-GNN · PeMS-LA, MCAR | MAE: 22.24 ± 0.25 -> 18.21 ± 0.12 | Helped |
| Add motif-based adjacency matrices (ablation) | Motif-GNN · Air Quality, MCAR | MAE: 18.16 ± 0.38 -> 15.43 ± 0.32 | Helped |
| Add motif-based adjacency matrices (ablation) | Motif-GNN · PeMS-LA, MCAR | MAE: 23.91 ± 0.25 -> 18.21 ± 0.12 | Helped |
| Use motif-wise multi-view attention instead of mean aggregation (ablation) | Motif-GNN · Air Quality, MCAR | MAE: 19.21 ± 0.26 -> 15.43 ± 0.32 | Helped |
| Use motif-wise multi-view attention instead of mean aggregation (ablation) | Motif-GNN · PeMS-LA, MCAR | MAE: 21.87 ± 0.29 -> 18.21 ± 0.12 | Helped |
Limitations
- Gains reported only against baselines on four datasets; no comparison with newer networked time series imputation methods beyond PoGeVon.
- Motif-based fusion adds computational cost, as the paper notes that separate GNNs per adjacency would be expensive, motivating fusion instead.
- Hardware, training cost, and hyperparameter details are not reported in the main text.
Experimental setup
Datasets: Air Quality (Yi et al. 2016) PeMS-LA PeMS-BA PeMS-SD
Benchmarks: Air Quality PeMS-LA PeMS-BA PeMS-SD
Metrics: MAE MRE (%)
Baselines: Mean Imputation Matrix Factorization GRU-D BRITS GRIN SPIN TimesNet PoGeVon
Builds on: GRIN SPIN Motif-based GNN (Lee et al. 2019) Chen et al. 2020 motif-induced adjacency
Closest papers by meaning
- Filling the G_ap_s: Multivariate Time Series Imputation by Graph Neural Networks (ICLR 2022 · similarity 0.92)
- Learning to Reconstruct Missing Data from Spatiotemporal Graphs with Sparse Observations (NeurIPS 2022 · similarity 0.88)
- Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data (AAAI 2024 · similarity 0.86)
- Graph Structure Learning for Spatial-Temporal Imputation: Adapting to Node and Feature Scales (AAAI 2025 · similarity 0.86)
- Discretizing Continuous Time Series for Imputation with Masked Diffusion Training (NeurIPS 2026 · similarity 0.85)
- Multi-Scale Dynamic Graph Learning for Time Series Anomaly Detection (Student Abstract) (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.