With the popularity of smart mobile devices, location-based services (LBS) have been widely applied. Predicting geographical locations from text holds significant value for smart cities and personalized travel. Existing research primarily focuses on the retrieval or prediction of labeled locations, such as cities or points of interest (POIs). However, in scenarios like autonomous driving navigation and autonomous logistics delivery, it is necessary to precisely predict the coordinates of unlabeled locations, for example, 200 meters northwest of a certain location. Consequently, we introduce a new task to infer fine-grained unlabeled locations from text. This task is particularly challenging because of the ambiguous text and the semantic gap between geographic and textual modalities. In this paper, we aim to construct an end-to-end fine-grained location prediction model to accurately predict the unlabeled locations mentioned in texts. First, we encode the geographic coordinates and transform the location prediction problem into a geographic encoding generation problem. Second, we propose a multi-scale cross-modal loss (MCL) to learn the implicit mapping between geographic and textual modalities. Lastly, we design a multi-task prediction model ULP to predict the coordinates of unlabeled locations. We conducted experiments on two real-world datasets, and the results show that our proposed method outperforms existing state-of-the-art retrieval-based methods.

The next point of interest (POI) recommendation task recommends POIs to users that they may be interested in next time based on their historical trajectories. This task holds value for both users and businesses. However, it has consistently faced the issue of cold-start caused by sparse user check-in data. Existing research mainly focuses on knowledge transfer among cities within the same data source, but these data are very rare. The abundance of available third-party data presents opportunities to improve cold-start performance, but it is not easy. This third-party data contain numerous entities, such as POIs and users, which have different representations and distributions across different data domains, making knowledge transfer difficult. To address these challenges, we propose the Tree-Based Cross-domain Knowledge Transfer (TCKT) model. First, we construct a multi-granularity Geographical Frequency Tree (GF-Tree), transforming the POI recommendation problem into a path generation problem. Second, we design a pre-training model to mine general user behavior patterns and spatio-temporal features among POIs from large-scale third-party data. Finally, we propose a dual-channel domain adaptation model to facilitate cross-domain knowledge transfer and improve cold-start performance. Experimental results on three public datasets demonstrate that our method outperforms state-of-the-art (SOTA) baseline methods.