摘要:
地理空间网络(geospatial network)是指网络中的节点或边的属性与地理位置相关的网络,能够同时表达地理研究中的地理实体及实体间的关系,是GIS研究中重要的一种数据类型。随着深度表征学习的发展,利用深度模型自动提取地理空间网络表征逐渐成为当前主流的范式。现有表征学习范式通过拟合观测数据中的相关性作为可学习的信号,但无法处理地理空间网络的某些特性带来的挑战。因果表征学习旨在为现有表征学习范式进一步引入因果知识作为学习约束,具有解决这类挑战的潜力。论述了为地理空间网络表征引入因果性视角的必要性,并对地理空间网络因果表征学习现有进展进行了总结。为了更好地提炼地理空间网络因果表征学习的重要概念及基本问题,以因果学习中的常用框架——结构因果模型(Structural Causal Model,SCM)为主线,将地理空间网络因果学习形式化为三个子任务。同时,构建了地理空间网络表征学习的统一框架,基于该框架总结了引入因果学习策略的意义。将地理空间网络中的因果关系分为空间因果和时间因果两类,并对现有研究分别进行了总结和梳理。在此基础上,对现有研究存在的挑战进行了总结,同时探讨了未来研究的方向。除此之外,还对通用的地理空间网络表征学习框架进行了抽象,为地理空间网络研究引入其他假设提供了可参考的骨架。
Abstract:
A geospatial network refers to a graph whose nodes or edges are associated with geographic locations. As an important data type in GIS, it simultaneously models geographical entities and their relationships. With the development of deep representation learning, leveraging deep models to automatically extract representations of geospatial networks has become the mainstream paradigm. The existing representation learning paradigm adopts the correlations observed in data as learnable signals but is unable to address the challenges posed by certain characteristics of geospatial networks. Causal representation learning aims to introduce causal knowledge as a learning constraint within the existing paradigm, offering potential solutions to such challenges. This article discusses the necessity of introducing causal assumptions for geospatial network representation and summarizes the current progress in causal representation learning for geospatial networks. To better distill the key concepts and fundamental issues of causal representation learning for geospatial networks, this article uses one of the most frequently used frameworks, the structural causal model, as the common thread, formalizing causal learning in geospatial networks into three sub-tasks. A unified framework for geospatial network representation learning is constructed, based on which the significance of introducing causal learning strategies is summarized. Additionally, this article categorizes causal relationships in geospatial networks into spatial causality and temporal causality, providing a comprehensive review and analysis of existing studies in each area. The challenges in current research are summarized, and future research directions are explored. Furthermore, this article abstracts a general framework for geospatial network representation learning, offering a reference skeleton for introducing other assumptions into geospatial network research.