Skip to content

FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts

Home / Papers / FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts

FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts

Authors: Yiji Zhao, Zihao Zhong, Ao Wang, Haomin Wen, Ming Jin, Yuxuan Liang, Huaiyu Wan, Hao Wu Date: 2026-04-20 Paper ID: openalex:2601.05174

Summary

FaST addresses the computational bottlenecks of existing spatial-temporal graph (STG) models by enabling efficient, large-scale, long-horizon forecasting. The framework utilizes an adaptive graph agent attention mechanism to replace expensive traditional graph convolutions and a novel parallel Mixture-of-Experts (MoE) structure based on Gated Linear Units. Experimental results demonstrate that FaST achieves superior long-horizon accuracy while significantly reducing computational overhead compared to current state-of-the-art baselines.

Key Contributions

  • Introduced FaST, an efficient MoE-based framework capable of 672-step long-horizon forecasting on large-scale graphs.
  • Developed Adaptive Graph Agent Attention to decouple computational complexity from graph scale.
  • Proposed a parallel MoE module utilizing Gated Linear Units (GLUs) for enhanced efficiency and scalability in STG forecasting.

Open Questions & Future Work

Key Concepts

  • adaptive-graph-agent-attention: An attention variant that uses graph agents to reduce computational complexity in large-scale spatial-temporal graph models.

Archivist Review

I approved the ‘Adaptive Graph Agent Attention’ as it represents a specific architectural innovation for the scalability-accuracy trade-off in STG forecasting. I also introduced a more precise open question regarding the integration of foundation models into STG frameworks, as the original submission was too generic and focused on future work plans rather than a technical bottleneck. The MoE module was rejected as it is a standard deep learning architectural choice rather than a specialized contribution to time-series forecasting.

Approved Concepts

  • Adaptive Graph Agent Attention: This mechanism addresses the core limitation of scaling STG forecasting to large networks by decoupling the graph’s complexity from the attention computation.

Approved Open Questions

  • Foundation Model STG Integration Bottleneck: As STG forecasting moves toward more complex, city-wide applications, foundational models offer a path to improved generalization, yet adapting them to maintain linear-time efficiency remains a challenging bottleneck.

Rejected Candidates

  • [concept] Parallel MoE Module (parallel-moe-module) - not_novel: Mixture-of-Experts combined with Gated Linear Units is a well-established architectural pattern in deep learning and does not warrant a novel standalone entry.

Metadata & Links

url
https://arxiv.org/abs/2601.05174
paper_id
2601.05174
paper_source
openalex
domain
time-series
tags
time-series-forecastinggraph-neural-networksmixture-of-experts
architectures
datasets
concept_slugs
adaptive-graph-agent-attention
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-23T05:47:48Z