High-dimensional urban dynamic patterns perception under the perspective of human activity semantics and spatiotemporal coupling

Abstract

As urbanization accelerates, megacities are emerging globally. Various human activities shape dynamic urban spaces, understanding dynamic performance implicit within them is essential for developing smart cities. Previous studies on urban dynamic patterns mainly focused on the spatiotemporal dimensions, unable to explain the joint effects of higher-dimensional patterns. In fact, large-scale social media data encapsulate human activity features across multiple dimensions, including semantics, space, and time, whose combined effects drive the formation of high-dimensional urban dynamic patterns. This study proposes a framework that expands the activity semantics dimension on top of spatiotemporal dimensions and perceive these patterns through high-dimensional feature coupling. Activity semantics are extracted from social media texts using ERNIE 3.0, a large-scale knowledge-enhanced pre-trained model. Data with three features dimensions are coupled into high-order tensors, and tensor decomposition uncovers key patterns. A case study using Weibo check-in records within Beijing’s Sixth Ring Road extracted ten distinct activity semantics, and interpretable patterns along each dimension. Through core tensors, we identified eight urban dynamic patterns under various states and their corresponding activity complexity changes. Additionally, correlations between activity semantics (dynamic attributes) and fixed facility configurations (static attributes) were explored using Point of Interest (POI) data. The results confirm the advantages of our method in exploring high-dimensional urban dynamic patterns.