Point-of-interest Representation Learning Model Towards Semantic Annotation

In recent years,the location-based social networking applications have generated a wealth of information-rich check-in data.People have conducted a lot of research on check-in data mining in the field of urban computing,which is important to boost production,facilitate residents’ lives,improve the ecological environment and efficiently use land resources.Among the various types of information related to check-in POIs,POI categories(e.g.,bars and museums)are crucial,because they reflect the semantic characteristics of POIs and are helpful to understand human activities and behaviors.Although POI categories are of great significance as key information in check-in data,POI categories suffer from severe missing problem,for example,up to 30%of POIs in Foursquare system lacking category labels.Some existing methods for POI semantic annotation often only construct features for classification,or represent POIs simply by mining check-in trajectories,and then treat the POI semantic annotation as a downstream task.These methods fail to fully exploit the contextual information in the check-in sequences,or do not capture the correlations among POIs through the known category information.To solve the above problems,this paper devises a Tree-guided Multi-task Embedding model(TME for short)to learn effective representations of POIs and categories for the semantic annotation.TME constructs POI co-occurrence PMI matrix and POI-category co-occurrence PMI matrix by modeling POI context and category context,and decomposes both matrices to learn representations of POIs and categories in the same latent feature space.In addition,TME models the known POI categories and a predefined category hierarchy to capture the relatedness among categories to enrich and enhance the POI and category representations.This paper evaluates TME over the task of POI semantic annotation on two check-in datasets.Experimental results show that the proposed TME significantly improves the performance on the POI semantic annotation task compared to several state-of-the-art baselines.