实验室两项研究成果被 KDD 2026 接收 Two Lab Papers Accepted by KDD 2026
近日，实验室两项研究成果 “HURST: Heterogeneity-Adaptive Urban Foundation Models for Spatiotemporal Prediction via Self-Partitional Mixture-of-Spatial-Experts” 和 “RePatch: Learning Entropy-Guided Patch Structures with Quantized Representations for Time Series Forecasting” 被 KDD 2026 接收。
HURST 聚焦城市时空预测中的空间异质性问题，探索面向不同空间区域特征进行自适应建模的城市基础模型。该工作通过自划分的空间专家混合机制，使模型能够更灵活地刻画复杂城市环境中不同区域的时空模式，为构建具有更强适应能力的城市时空基础模型提供了新的思路。
RePatch 聚焦时间序列预测中的序列结构建模问题，从数据自身的信息特征出发，探索更加灵活的 Patch 构建方式。该工作结合熵引导的 Patch 结构学习与量化表示，使模型能够根据时间序列的内在特征学习更合适的表示结构，为时间序列预测提供新的建模方法。
链接：HURST: Heterogeneity-Adaptive Urban Foundation Models for Spatiotemporal Prediction via Self-Partitional Mixture-of-Spatial-Experts；RePatch: Learning Entropy-Guided Patch Structures with Quantized Representations for Time Series Forecasting
Recently, two research works from our lab, “HURST: Heterogeneity-Adaptive Urban Foundation Models for Spatiotemporal Prediction via Self-Partitional Mixture-of-Spatial-Experts” and “RePatch: Learning Entropy-Guided Patch Structures with Quantized Representations for Time Series Forecasting,” were accepted by KDD 2026.
HURST focuses on spatial heterogeneity in urban spatiotemporal prediction and explores adaptive modeling for different spatial regions within urban environments. By introducing a self-partitional mixture-of-spatial-experts mechanism, the work enables more flexible modeling of diverse spatiotemporal patterns across urban areas, providing a new direction for building more adaptive urban foundation models.
RePatch focuses on structural modeling for time series forecasting and explores a more flexible way of constructing patches based on the intrinsic information characteristics of time series. By combining entropy-guided patch structure learning with quantized representations, the method allows the model to learn more suitable representation structures from the data itself, offering a new approach to time series forecasting.
Links: HURST: Heterogeneity-Adaptive Urban Foundation Models for Spatiotemporal Prediction via Self-Partitional Mixture-of-Spatial-Experts; RePatch: Learning Entropy-Guided Patch Structures with Quantized Representations for Time Series Forecasting