JiuTian Chuanliu: A Large Spatiotemporal Model for General-purpose Dynamic Urban Sensing

TL;DR

This work presents GDHME, a large-scale spatiotemporal framework that unifies human mobility into a continuous-time people-region dynamic graph and learns general-purpose embeddings via self-supervised pretraining. The encoder maintains dynamic and static memories for both people and regions, while a specialized decoder enables next-location prediction with hard negative sampling; after pretraining, the embeddings support diverse downstream tasks across region- and people-centric applications. A comprehensive multi-task urban sensing benchmark and a real-world Beijing dataset evaluate the approach, demonstrating robust performance gains and the ability to deploy a JiuTian ChuanLiu Big Model for industry use. The method offers label-efficient, privacy-preserving insights suitable for smart city planning, marketing, and mobility services, with potential for expansion through heterogeneous data fusion in future work.

Abstract

As a window for urban sensing, human mobility contains rich spatiotemporal information that reflects both residents' behavior preferences and the functions of urban areas. The analysis of human mobility has attracted the attention of many researchers. However, existing methods often address specific tasks from a particular perspective, leading to insufficient modeling of human mobility and limited applicability of the learned knowledge in various downstream applications. To address these challenges, this paper proposes to push massive amounts of human mobility data into a spatiotemporal model, discover latent semantics behind mobility behavior and support various urban sensing tasks. Specifically, a large-scale and widely covering human mobility data is collected through the ubiquitous base station system and a framework named General-purpose and Dynamic Human Mobility Embedding (GDHME) for urban sensing is introduced. The framework follows the self-supervised learning idea and contains two major stages. In stage 1, GDHME treats people and regions as nodes within a dynamic graph, unifying human mobility data as people-region-time interactions. An encoder operating in continuous-time dynamically computes evolving node representations, capturing dynamic states for both people and regions. Moreover, an autoregressive self-supervised task is specially designed to guide the learning of the general-purpose node embeddings. In stage 2, these representations are utilized to support various tasks. To evaluate the effectiveness of our GDHME framework, we further construct a multi-task urban sensing benchmark. Offline experiments demonstrate GDHME's ability to automatically learn valuable node features from vast amounts of data. Furthermore, our framework is used to deploy the JiuTian ChuanLiu Big Model, a system that has been presented at the 2023 China Mobile Worldwide Partner Conference.

Figures (13)

• Figure 1: The overview of JiuTian·Chuanliu.
• Figure 2: The overall pipeline of GDHME. The continuous-time human mobility encoder maintains memories for each region node and each people node. When people appear at new locations, the memories are updated according to interactions between people and regions. Then, the node representations are generated with the updated memories by the embedding module. Meanwhile, a continuous time human mobility decoder is specially designed to predict next location of the people in stage 1. In stage 2, the well-learned node representations will be utilized to solve various downstream tasks.
• Figure 3: Illustration of the continuous-time human mobility encoder.
• Figure 4: Illustration of the continuous-time human mobility decoder.
• Figure 5: Illustration of the regions.