Learning context-aware region similarity with effective spatial normalization over Point-of-Interest data

With the increasing availability of Point -of -Interest (PoI) data driven by the widespread adoption of location -based services, there is a growing demand to comprehend the similarities among different regions. Existing works generally regard regions with the similar number or distribution of PoI as similar. However, in practice, the region similarity depends on the surrounding environment as well, owing to the inherent imbalance observed in urban PoI data. In addition, the existing literature does not consider different task contexts when applying regional similarity. This paper primarily investigates how to appropriately represent the distribution of PoIs to measure the similarity between regions for diverse spatial and application contexts. A Context -Aware REgion similarity learning framework (CARE) is proposed, which utilizes a novel spatial normalization technique to capture the unique role of each region within its neighborhoods in order to address the inherent imbalance issue in urban data and a self -supervised contrastive learning method with triplet loss to tackle problem of missing labeled data in datasets. Considering the variations in how different applications define region similarity, CARE prompts large language model to adjust learned region embeddings and adapt them for specific application requirements. We conduct experiments on three datasets and two downstream tasks. Results show that compared with the baseline models, our model improves the performance of site selection recommendation and crime prediction by up to 19.60% and 35.86%, respectively.