Research map / Graph learning / Knowledge and temporal graphs

ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Temporal graph neural architectures · 5 citations (OpenAlex)

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 ST-FiT, a plug-and-play framework that combines VAE-based temporal mix-up augmentation with Gumbel-Softmax spatial topology learning so STGNNs can forecast on nodes lacking training temporal data without fine-tuning.

Problem

Spatial-temporal forecasting when most nodes have no temporal data during training, requiring inductive generalization to those nodes without costly fine-tuning.

Approach

Learns a VAE latent manifold of available temporal data, performs mix-up between latent codes to generate new temporal dependencies, and learns a sparse spatial topology via Gumbel-Softmax with a threshold, iteratively trained with an STGNN backbone.

Key results

• On PEMS04 ST-FiT reaches MAE 25.11 / RMSE 39.30 / MAPE 17.23, outperforming TransGTR (32.76 / 48.94 / 26.87) by about 27% MAE without fine-tuning.
• ST-FiT exceeds its STGCN backbone by up to 40.0% in MAE, 38.6% in RMSE and 55.6% in MAPE on nodes without training data.
• On PEMS03 ST-FiT (18.40 MAE) is behind fine-tuned TransGTR (17.50 MAE), and on PEMS08 ST-FiT (25.09 MAE) is well behind TransGTR (18.00 MAE).

Measured effects reported in the paper

Surprising findings

• Sparser learned topologies generally perform better, with performance peaking at threshold 0.9 and degrading when overly sparse.
• On PEMS08 ST-FiT trails the fine-tuned TransGTR substantially (25.09 vs 18.00 MAE), a gap the paper attributes to TransGTR's use of abundant temporal data during fine-tuning.

Limitations

• Inductive setting only evaluated on three PEMS traffic datasets with a single backbone (STGCN) in main results.
• Mix-up in latent space does not rigorously guarantee generated samples lie on the learned manifold.
• Training time per epoch is roughly double that of STGCN due to two forward passes per iteration.

Experimental setup

Datasets: PEMS03 PEMS04 PEMS08

Benchmarks: PEMS03 PEMS04 PEMS08

Metrics: MAE RMSE MAPE

Baselines: Historical Average FC-LSTM STGCN STGODE TransGTR

Builds on: STGCN VAE Gumbel-Softmax TransGTR

Closest papers by meaning

• STD-PLM: Understanding Both Spatial and Temporal Properties of Spatial-Temporal Data with PLM (AAAI 2025 · similarity 0.89)
• Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling (ICML 2024 · similarity 0.88)
• Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting (ICML 2025 · similarity 0.88)
• Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data (AAAI 2024 · similarity 0.88)
• How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning (NeurIPS 2025 · similarity 0.88)
• Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting (ICLR 2025 · similarity 0.88)