AI News

The paper introduces IPGeoAI, a deep learning architecture for city-level IP geolocation that treats the task as sequential modeling rather than static address lookup. It uses a Transformer Encoder to model hierarchical dependencies in IP subnet structures, while a zero-shot LLM pipeline converts noisy Autonomous Systems descriptions into structured metadata such as organization type and geographic scope. A Multi-Head Cross-Attention module fuses those semantic signals with numerical network-topology features to address geographic ambiguity. In offline evaluation on a proprietary dataset spanning 200,000 cities, the authors report that IPGeoAI outperformed a leading external vendor at city-level granularity, improved city-level accuracy by 6%, and extended coverage to 100% of traffic through hierarchical inference that first refines country-level signals. Large-scale online production tests also showed a statistically significant 0.35% improvement in the metric for first-tier downstream use cases. The reported results concern offline and production evaluations; the abstract does not provide further dataset or experimental details.