Urban In-context Learning: A New Paradigm for Urban Indicator Prediction

Recent years have witnessed the rapid development of urbanization. Specifically, urban indicator prediction has become an important tool for urban planning and decision-making and promoting the process of urbanization. However, the existing methods have the following two drawbacks. First, they follow the ''pre-training and fine-tuning'' paradigm, which is time-consuming and resource-intensive. Second, to encode urban knowledge for downstream tasks effectively, complex pre-training tasks must be designed to train the model in a task-agnostic manner while ensuring generalization. In this work, we propose UrbanICL, an urban in-context learning framework as a new paradigm for urban indicator prediction. Compared to directly predicting urban indicators, we obtain predictions for new regions by aggregating the downstream labels of similar regions. Specifically, a retrieval-based urban in-context learning module is proposed to retrieve regions with similar urban semantics and aggregate their corresponding labels to make predictions for new regions. We also design a region-dependent distribution learning module to learn the new distribution of unknown regions and facilitate the adaptation of UrbanICL for distributional shifts and outliers. Our framework, with in-context learning, brings a new insight for urban indicator prediction. We conduct extensive experiments on real-world datasets collected from three cities. The experiment results demonstrate the effectiveness of UrbanICL, even in an extremely low-consumption and time-efficient manner.