Research map / Probabilistic and scientific modeling / Molecules and proteins
Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach
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 E²-CSTP, a multi-modal spatio-temporal forecasting framework combining cross-modal attention and gating, a dual-branch backdoor-style causal adjustment, and a GCN+Mamba encoder for efficiency, evaluated on four datasets.
Problem
Spatio-temporal prediction struggles with inadequate fusion of heterogeneous multi-modal data, confounding factors obscuring causal relations, and the quadratic cost of Transformer-based forecasters.
Approach
Fuse spatio-temporal, text and image modalities via cross-modal attention and gating; add a dual-branch training scheme with a latent-confounder intervention on the fused features; and replace attention with GCN (spatial) plus Mamba (temporal) for linear-cost spatio-temporal encoding.
Key results
- E²-CSTP achieves MAE 2.43 on Terra vs 2.47 for the best baseline (GPT-ST/UniST), and 2.99 on BikeNYC vs 3.31 for UniST.
- Reports up to 9.66% accuracy improvement and 17.37%–56.11% reduction in computational overhead versus Transformer-based prediction module alternatives on Terra.
- Per-epoch runtime on multi-modal datasets is slower than the strongest baseline, per the appendix; the speed-up is mainly on single-modal BikeNYC (about 20%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add text inputs (ablation) | E²-CSTP · Datasets with text modalities | prediction accuracy | Helped |
| Add image inputs (ablation) | E²-CSTP · Datasets with image modalities | prediction accuracy | Helped |
| Add DeepSHAP-based causal region identification (ablation) | E²-CSTP · Spatio-temporal prediction | prediction accuracy | Helped |
| Add causal intervention (ablation) | E²-CSTP · Confounding conditions | robustness | Helped |
| Add GCN spatial encoding (ablation) | E²-CSTP · Spatio-temporal prediction | prediction accuracy | Helped |
| Add Mamba temporal encoding (ablation) | E²-CSTP · Spatio-temporal prediction | prediction accuracy | Helped |
| Replace Transformer-based prediction modules with GCN-Mamba STED | E²-CSTP · Terra | prediction accuracy: 1.78%–5.45% | Helped |
| Replace Transformer-based prediction modules with GCN-Mamba STED | E²-CSTP · Terra | computational overhead: 17.37%–56.11% | Helped |
| Replace Transformer-based prediction modules with GCN-Mamba STED | E²-CSTP · BjTT, GreenEarthNet, and BikeNYC | MAE: 15.64% to 44.20% | Helped |
| Replace Transformer-based prediction modules with GCN-Mamba STED | E²-CSTP · BjTT, GreenEarthNet, and BikeNYC | per-epoch runtime: 14.20% to 89.25% | Helped |
| Use the proposed model rather than the second-best method | E²-CSTP · Terra, BjTT, GreenEarthNet, and BikeNYC | MAE: 1.61% to 9.66% | Helped |
| Use the proposed model rather than the strongest baseline | E²-CSTP · Multi-modal datasets | per-epoch runtime | Hurt |
Surprising findings
- The appendix reports E²-CSTP per-epoch runtime is slower than the strongest baseline on multi-modal datasets, contrary to the headline efficiency claim.
Limitations
- Only spatio-temporal sequence, text and image modalities are handled; audio, LiDAR, and social signals are not considered.
- Multi-modal datasets show slower per-epoch runtime than the strongest baseline, per the appendix.
- Causal adjustment relies on latent confounders approximated by learned variables, so causal validity is heuristic.
Experimental setup
Datasets: Terra BjTT GreenEarthNet BikeNYC
Benchmarks: Terra BjTT GreenEarthNet BikeNYC
Metrics: MAE RMSE MAPE training time per-epoch runtime
Baselines: D2STGNN ST-SSL HimNet NuwaDynamics CaPaint GPT-ST UniST T3 From News to Forecast (FNF)
Builds on: Mamba DeepSHAP backdoor adjustment BERT
Closest papers by meaning
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.90)
- GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding (NeurIPS 2025 · similarity 0.88)
- GeoPro-Net: Learning Interpretable Spatiotemporal Prediction Models Through Statistically-Guided Geo-Prototyping (AAAI 2025 · similarity 0.88)
- How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning (NeurIPS 2025 · similarity 0.87)
- Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting (AAAI 2025 · similarity 0.87)
- Discovering Latent Causal Graphs from Spatiotemporal Data (ICML 2025 · similarity 0.87)
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.