Research map / Time series / Causal and generative time series
GraFITi: Graphs for Forecasting Irregularly Sampled Time Series
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 GraFITi, which represents irregularly sampled multivariate time series with missing values as a sparse bipartite Sparsity Structure Graph and forecasts by predicting target edge weights with an attention-based GNN.
Problem
Forecasting irregularly sampled multivariate time series with missing values (IMTS), where ODE-based models are slow and need missing-value indicators as extra channels.
Approach
Converts the IMTS into a bipartite graph of channel and timepoint nodes with observations as edges, reformulates forecasting as edge weight prediction for query edges, and uses attention-based GNN layers that update node and edge embeddings without missing-value indicators.
Key results
- GraFITi achieves MSE 0.225 on MIMIC-IV forecasting next three steps, a 17.2% improvement over the next best LinODEnet (0.272).
- GraFITi is over 5 times faster than LinODEnet on MIMIC-IV and USHCN, and about twice as fast on Physionet'12 and MIMIC-III.
- GraFITi improves over LinODEnet by 11.2% on MIMIC-III and 4.3% on Physionet'12 for the three-step forecast.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GraFITi instead of LinODEnet | IMTS forecasting · USHCN | MSE: 0.300 ± 0.060 -> 0.272 ± 0.047 | Helped |
| Use GraFITi instead of LinODEnet | IMTS forecasting · MIMIC-III | MSE: 0.446 ± 0.033 -> 0.396 ± 0.030 | Helped |
| Use GraFITi instead of LinODEnet | IMTS forecasting · MIMIC-IV | MSE: 0.272 ± 0.002 -> 0.225 ± 0.001 | Helped |
| Use GraFITi instead of LinODEnet | IMTS forecasting · Physionet’12 | MSE: 0.299 ± 0.001 -> 0.286 ± 0.001 | Helped |
| Use GraFITi instead of LinODEnet | IMTS forecasting · MIMIC-III, varying observation and forecast ranges | MSE: 8.5% | Helped |
| Use GraFITi instead of LinODEnet | IMTS forecasting · MIMIC-IV, varying observation and forecast ranges | MSE: 15.5% | Helped |
| Use GraFITi instead of LinODEnet | IMTS forecasting · Physionet’12, varying observation and forecast ranges | MSE: 2.6% | Helped |
| Use GraFITi instead of the fastest ODE model | IMTS forecasting · MIMIC-IV and USHCN | runtime: over 5 times faster | Helped |
| Increase sparsity in the asynchronous time-series input | GraFITi · MIMIC-III, AsTS with varying sparsity | MSE: 0.931 -> 0.413 | Hurt |
Surprising findings
- Standard MTS forecasting models adapted to IMTS with missing-value indicators perform worse than dedicated IMTS forecasting models.
- Performance degrades as sparsity increases and the graph becomes disconnected in asynchronous settings, where GraFITi underperforms LinODEnet at high sparsity.
Limitations
- Performance deteriorates with increasing sparsity; in asynchronous time series the graph becomes disconnected and GraFITi underperforms LinODEnet.
- Cannot handle metadata associated with the IMTS.
- High variance on the USHCN dataset makes comparisons difficult.
Experimental setup
Datasets: Physionet'12 MIMIC-III MIMIC-IV USHCN synthetic asynchronous time series derived from MIMIC-III
Benchmarks: USHCN MIMIC-III MIMIC-IV Physionet'12
Metrics: MSE run time
Baselines: GRU-ODE-Bayes Neural Flows CRU LinODEnet mTAN Informer+ FedFormer+ DLinear+ NLinear+ GRU-D
Builds on: Graph Attention Network Transformer multi-head attention mTAN LinODEnet
Closest papers by meaning
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.86)
- Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling (ICML 2024 · similarity 0.85)
- HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting (ICML 2025 · similarity 0.85)
- Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs (ICLR 2025 · similarity 0.85)
- Probabilistic Forecasting of Irregularly Sampled Time Series with Missing Values via Conditional Normalizing Flows (AAAI 2025 · similarity 0.84)
- Motif-aware Graph Neural Networks for Networked Time Series Imputation (AAAI 2025 · similarity 0.84)
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.