Research map / Graph learning / Knowledge and temporal graphs

KITS: Inductive Spatio-Temporal Kriging with Increment Training Strategy

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Continuous-time dynamic graph learning · 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 KITS, an inductive spatio-temporal kriging method that replaces masking-based decrement training with an increment strategy inserting virtual nodes, plus reference-based feature fusion and node-aware cycle regulation.

Problem

Inductive spatio-temporal kriging (inferring values at unobserved sensor nodes from observed ones) suffers from a "graph gap" because training graphs built from masked observed nodes are much sparser than inference graphs including unobserved nodes.

Approach

Instead of masking observed nodes, insert empty virtual nodes into the training graph (Increment training strategy) to mimic unobserved nodes, then fuse each virtual node's features with its most similar observed node (RFF) and regulate virtual nodes with inverse-mask pseudo-label cycle training (NCR).

Key results

• KITS reduces MAE by 18.33% over the best inductive baseline on AQI and MAPE by 30.54% on NREL-MD (inductive setting, alpha=50%).
• On the transductive setting, KITS improves MAE by 20.61% on PEMS07 and MAPE by up to 45.50% on NREL-MD over GRIN and MPGRU.
• Ablation on METR-LA: the Increment strategy alone improves MAE by 5.69% over the Decrement strategy; full KITS reaches MAE 6.1276 vs 6.7597 for the decrement baseline.

Measured effects reported in the paper

Surprising findings

• Virtual graphs created by the increment strategy at missing ratios above 76% outperform the real transductive graph with unobserved nodes on METR-LA.
• Reported degree statistics show the decrement training graphs' average largest degree is over 40% lower than the inference graph's, while increment graphs come close.

Limitations

• Assumes a rough estimate of the missing ratio alpha is available at training time
• Results reported mainly at a fixed 50% missing ratio with a fixed random seed for the main tables
• Some baselines required adaptation (validation on valid set) and some cannot run on SEA-LOOP and PEMS07 due to missing coordinates

Experimental setup

Datasets: METR-LA PEMS-BAY SEA-LOOP PEMS07 AQI-36 AQI NREL-AL NREL-MD

Benchmarks: METR-LA PEMS-BAY SEA-LOOP PEMS07 AQI-36 AQI NREL-AL NREL-MD

Metrics: MAE MAPE MRE RMSE R2

Baselines: Mean imputation OKriging KNN KCN IGNNK LSJSTN INCREASE GLTL MPGRU GRIN

Builds on: IGNNK GraphSAGE GRIN

Closest papers by meaning

• Graph Structure Learning for Spatial-Temporal Imputation: Adapting to Node and Feature Scales (AAAI 2025 · similarity 0.84)
• Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling (ICML 2024 · similarity 0.84)
• Learning from Highly Sparse Spatio-temporal Data (NeurIPS 2024 · similarity 0.84)
• Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.83)
• RDPI: A Refine Diffusion Probability Generation Method for Spatiotemporal Data Imputation (AAAI 2025 · similarity 0.83)
• Efficient Unlearning for Spatio-temporal Graph (Student Abstract) (AAAI 2025 · similarity 0.83)