Abstract
In this paper, we present an approach for effectively classifying Points of Interest (POIs) that are represented only by their name and location (coordinates). Most existing approaches make the assumption that the handled POIs carry a wealth of metadata (e.g. reviews, ratings, working hours, price ranges). Consequently, such methods rely on semantically rich POI profiles and exploit them to develop correspondingly rich, and thus more accurate, POI classification models. However, in several real world scenarios, assuming the existence of such rich POI profiles is unrealistic. Contrary to existing works, we propose a method that can produce accurate category recommendations based only on the minimum amount of initially available POI metadata (name, coordinates) combined with open and straightforwardly accessible metadata drawn from OpenStreetMap. To this end, we propose a set of textual and neighbourhood-based training features, capturing POI properties as well as their relations with their spatial neighborhoods. These features are fed into several classification algorithms and are evaluated on a proprietary POI dataset of a geo-marketing company and the Yelp POI dataset.
Nikos Kostagiolas
CS Ph.D. Student
Hi! I am an Ph.D. student in Computer Science at the CaSToRC unit of the Cyprus Institute under the supervision of Prof. Mihalis Nicolaou, focusing on Representation Learning and Computer Vision.