Generating evolving region embedding with memory-based graph for dynamic urban sensing

Recently, learning region embeddings from mobility data has become a prevalent solution to various smart city tasks. However, existing methods primarily focus on learning static embeddings, which often fail to capture temporal dynamics critical for tasks like travel time estimation and dynamic crime prediction. To fill the gap, this paper introduces an Evolving Urban Region Embedding method (EvolveURE), a generic and evolving region embedding framework tailored for dynamic urban sensing tasks. EvolveURE introduces a memory-based embedding module that continuously updates region embeddings by encoding time-aware region-region interactions and recursively updating region memories. Additionally, a time-aware mobility graph reconstruction task is designed to capture detailed temporal information by reconstructing both the magnitude and temporal density of mobility data. To enhance embedding stability and generalization, a cross-scale embedding reconstruction task is proposed, leveraging insights from different data scales and parameter updating mechanisms. The learned region embeddings are tested on diverse downstream urban sensing tasks. Experimental results demonstrate the proposed approach achieves superior results on dynamic urban sensing tasks thanks to these time-aware designs. Codes have been released on https://github.com/rabbityi1999/EvolveURE/.