Research map / Video / Video segmentation and tracking
Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework
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 CMuST, a continuous multi-task spatiotemporal learning framework with a multi-dimensional interaction network (MSTI) and a rolling adaptation scheme (RoAda) that uses autoencoder task prompts and weight-variance-based freezing, evaluated on three new urban datasets (NYC, SIP, Chicago).
Problem
Urban spatiotemporal models are trained per task and fail to generalize to new domains or cold-start tasks, and existing multi-task spatiotemporal methods do not explicitly capture task-level commonality and personalization.
Approach
Combine cross- and self-attention across spatial, temporal and observation dimensions (MSTI) with a rolling training scheme that freezes low-variance weights across sequential tasks to capture shared patterns, plus autoencoder-based task summaries as prompts for task-specific refinement.
Key results
- CMuST achieves best MAE on most of the 9 tasks, e.g., NYC Crowd In MAE 11.1533 vs. best baseline PromptST 11.0036 (CMuST is second there) and NYC Crowd Out 12.9088 vs. PromptST 13.0237.
- Under data sparsity (25% nodes) on NYC Crowd In, CMuST reaches MAE 12.1611 vs. PromptST 12.8362.
- Multi-task methods improve over single-task methods by an average of 8.51% per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spatial-context and temporal-context cross-interactions (ablation) | CMuST MSTI · Chicago | prediction performance | Helped |
| Add stable-weight separation and freezing during rolling adaptation (ablation) | CMuST RoAda · Chicago | prediction performance | Helped |
| Add task-specific prompts to preserve task diversity (ablation) | CMuST RoAda · Chicago | prediction performance | Helped |
| Use CMuST instead of single-task baselines | NYC, SIP, and Chicago prediction tasks · NYC, SIP, and Chicago | MAE | Mixed |
| Use CMuST instead of PromptST | NYC Crowd In · NYC Crowd In | MAE: 11.0036 vs 11.1533 | Hurt |
| Use CMuST for prediction with reduced spatial nodes or temporal samples | NYC Crowd In · NYC Crowd In data-scarcity scenarios | MAE | Helped |
| Add more tasks to the learning process | CMuST · NYC individual tasks | prediction performance | Helped |
| Use cold-start adaptation instead of training a task alone | CMuST · NYC, SIP, and Chicago tasks | MAE | Helped |
| Increase task-prompt dimension up to 72 (ablation) | CMuST · three city datasets | MAE | Helped |
| Increase task-prompt dimension beyond 72 (ablation) | CMuST · three city datasets | MAE | Hurt |
| Increase MSTI attention heads up to four (ablation) | CMuST · three city datasets | prediction performance | Helped |
| Increase MSTI attention heads beyond four (ablation) | CMuST · three city datasets | prediction performance | Hurt |
Limitations
- Confined to a single urban system and only transportation-related tasks.
- Experiments only with regression tasks.
Experimental setup
Datasets: NYC taxi trip data NYC crowd flow SIP traffic flow and speed Chicago taxi trip data Chicago traffic accident risk data
Benchmarks: NYC SIP Chicago
Metrics: MAE MAPE
Baselines: DCRNN AGCRN GWNET STGCN MTGNN STEP PromptST GMAN ASTGCN STTN UniST UniTime
Builds on: PromptST STEP MTGNN GMAN UniST
Closest papers by meaning
- Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.91)
- Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting (AAAI 2025 · similarity 0.87)
- GeoPro-Net: Learning Interpretable Spatiotemporal Prediction Models Through Statistically-Guided Geo-Prototyping (AAAI 2025 · similarity 0.87)
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.86)
- Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation (ICLR 2024 · similarity 0.85)
- UniSTD: Towards Unified Spatio-Temporal Learning across Diverse Disciplines (CVPR 2025 · 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.