CAVE-SC: Inferring categories for venues using check-ins

The categories of venues are labels, such as Museum and Bar, assigned to geographic locations. They are often incomplete and new venues may constantly arise, making automatic venue labeling essential. In this study, we proposed a category-aware venue embedding model with similarity constraints (CAVE-SC), which is an automatic feature engineering method to infer category labels for venues. CAVE-SC first models the co-occurrence relations between a given venue and its two types of sequential contexts (namely, the venue context and the category context) based on a softmax loss and then projects both venues and categories into the same latent space. Furthermore, CAVE-SC models the similarity constraints posed by the relationships between venues and their categories based on a newly designed triplet loss, assuming that the distance between a venue and its category is smaller than the distance between the venue and a randomly sampled category in the embedding space. Finally, linear interpolation is used to balance the two losses, thereby generating discriminative feature representations of venues. Venue labeling can be performed based on the distances between venue representations and category representations. Extensive experiments on two check-in datasets demonstrate the superiority of CAVE-SC over the baseline methods.