ICLR 2025 | 时空数据(Spatial-Temporal)高分论文总结

时空探索之旅

发布于 2024-12-19 16:43:21

2700

文章被收录于专栏：时空探索之旅时空探索之旅

ICLR2025已经结束了讨论阶段，进入了meta-review阶段，分数应该不会有太大的变化了，本文总结了其中时空数据(Spatial-Temporal)高分的论文。如有疏漏，欢迎大家补充。

挑选原则：均分要大于等于6（≥6，即使有3，但是有8或者更高的分拉回来也算）

时空数据Topic：时空预测，轨迹生成，LLM以及基础模型等内容。总计10篇

点击文末阅读原文跳转笔者知乎链接（跳转论文链接更方便）。

WardropNet: Traffic Flow Predictions via Equilibrium-Augmented Learning
CityBench: Evaluating the Capabilities of Large Language Models for Urban Tasks
Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting
Learning Spatiotemporal Dynamical Systems from Point Process Observations
High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation
PIMRL: Physics-Informed Multi-Scale Recurrent Learning for Spatiotemporal Prediction
STOP! A Out-of-Distribution Processor with Robust Spatiotemporal Interaction
Deep Random Features for Scalable Interpolation of Spatiotemporal Data
Does Vector Quantization Fail in Spatio-Temporal Forecasting? Exploring a Differentiable Sparse Soft-Vector Quantization Approach
DiffMove: Human Trajectory Recovery via Conditional Diffusion Model

1 WardropNet: Traffic Flow Predictions via Equilibrium-Augmented Learning

链接：https://openreview.net/forum?id=7FHSPd3SRE

分数：8655

关键词：交通平衡

keywords： structured learning, combinatorial optimization augmented machine learning, traffic equilibrium prediction

2 CityBench: Evaluating the Capabilities of Large Language Models for Urban Tasks

链接：https://openreview.net/forum?id=oIWN7eMhTb

分数：866510

关键词：大模型评估，城市模拟

keywords：LLM，benchmark

TL; DR：We propose a simulator based global scale benchmark to evaluate the performance of large language models on various urban tasks.

3 Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting

链接：https://openreview.net/forum?id=FRzCIlkM7I

分数：38888

关键词：持续学习，时空预测

keywords：Spatio-temporal Graph, Continual Forecasting, Tuning Principle

TL; DR：We introduce EAC, which follows the two fundamental tuning principle and learns prompt parameters pool only through expand and compress, simply, effectively and efficiently solving the continual spatio-temporal graph forecasting problem.

EAC

4 Learning Spatiotemporal Dynamical Systems from Point Process Observations

链接：https://openreview.net/forum?id=37EXtKCOkn

分数：8688

关键词：点过程，时空动力学模型

keywords： dynamics, spatiotemporal, neural, PDE, ODE

5 High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation

链接：https://openreview.net/forum?id=Cjz9Xhm7sI

分数：888

关键词：气象预测，Mamba

keywords： 3D Gaussian, Dynamic Reconstruction, Radar Prediction, Weather Nowcasting

6 PIMRL: Physics-Informed Multi-Scale Recurrent Learning for Spatiotemporal Prediction

链接：https://openreview.net/forum?id=w3rbBVJ9Jg

分数：8566

关键词：微分方程，时空预测

keywords：PDEs, physics encoding, data-driven modeling

TL; DR：We introduce a new multi-scale framework calling physics-informed multi-scale recurrent learning (PIMRL) framework to effectively utilize multi-scale time data.

7 STOP! A Out-of-Distribution Processor with Robust Spatiotemporal Interaction

链接：https://openreview.net/forum?id=85WHuB5CUK

分数：6666

关键词：分布外泛化，时空预测

keywords：Spatiotemporal learning; out-of-distribution learning; spatiotemporal prediction

8 Deep Random Features for Scalable Interpolation of Spatiotemporal Data

链接：https://openreview.net/forum?id=OD1MV7vf41

分数：388

关键词：地球科学，高斯过程

keywords：Random Features, Deep Gaussian Processes, Bayesian Deep Learning, Remote Sensing

TL; DR：We propose a scalable Bayesian deep learning framework to interpolate remote sensing data for increased accuracy and flexibility.

9 Does Vector Quantization Fail in Spatio-Temporal Forecasting? Exploring a Differentiable Sparse Soft-Vector Quantization Approach

链接：https://openreview.net/forum?id=4CFVPCYfJ9

分数：5856

关键词：向量量化，稀疏回归

keywords：spatio-temporal forecasting, vector quantilization, sparse regression, differentiable, soft

TL; DR：Does Vector Quantization Fail in Spatio-Temporal Forecasting? Exploring a Differentiable Sparse Soft-Vector Quantization Approach

10 DiffMove: Human Trajectory Recovery via Conditional Diffusion Model

链接：https://openreview.net/forum?id=VRFotuGLfM

分数：66865

关键词：轨迹恢复，扩散模型

keywords： Trajectory recovery, Diffusion model, Self-supervised learning, Human mobility

TL; DR：This paper presents DiffMove, a novel conditional diffusion based method for recovering human trajectories from incomplete data, outperforming existing approaches by an average of 11% in recall.

DiffMove