Enabling the interpretability of pretrained venue representations using semantic categories

Abstract The growing popularity of location-based social networks gives rise to a tremendous amount of social check-ins data, which are broadly used in previous studies to produce dense venue representations for various trajectory mining tasks. In this work, we focus on the interpretability of venue representations, an essential property that existing methods fail to provide. We propose two novel models to generate interpretable and easy-to-understand venue representations. The first model, CEM, is a category-aware (a category may be a restaurant, a mall, etc.) check-in embedding model and generates venue and category representations by capturing the sequential patterns of check-in records. With the second model, XEM, each dimension of the venue representation corresponds to a semantic anchor (i.e., a category) and can be interpreted as a coherent topic. We conduct extensive experiments using real-world check-in datasets for venue similarity computation and venue semantic annotation, and empirically show that introducing interpretability to the venue representations improves the performance of various downstream tasks.