NeighborGeo: IP geolocation based on neighbors

Abstract

IP geolocation is crucial in fields such as cybersecurity, e-commerce, and social media. Current mainstream graph neural network methods have advanced localization accuracy by reframing the IP geolocation task as a node regression problem within an attribute graph, leveraging features to model the connectivity between nodes. However, in practical applications, landmarks are often scattered, irregular, and susceptible to outliers, which limits their accuracy due to the unreliability of landmark selection and relationship learning. To address these challenges, this paper introduces a novel IP geolocation model based on graph structure learning, termed NeighborGeo. This model employs reparameterization and supervised contrastive learning to precisely capture and selectively reinforce specific neighbor relationships between nodes in order to optimize structural representations. By accurately capturing and utilizing neighbors, this model achieves accurate predictions. Experimental results demonstrate that, on open-source datasets from New York, Los Angeles, and Shanghai, NeighborGeo achieves significantly higher localization accuracy compared to existing methods, particularly in scenarios with unevenly distributed landmarks.