Abstract
IP Geolocation databases are widely used in online services to map end userIP addresses to their geographical locations. However, they use proprietarygeolocation methods and in some cases they have poor accuracy. We propose asystematic approach to use publicly accessible reverse DNS hostnames forgeolocating IP addresses. Our method is designed to be combined with othergeolocation data sources. We cast the task as a machine learning problem wherefor a given hostname, we generate and rank a list of potential locationcandidates. We evaluate our approach against three state of the art academicbaselines and two state of the art commercial IP geolocation databases. We showthat our work significantly outperforms the academic baselines, and iscomplementary and competitive with commercial databases. To aidreproducibility, we open source our entire approach.