Accurate city‑level IP geolocation underpins many modern digital services, from localized content delivery to rights enforcement. Conventional heuristic or database lookup approaches struggle with the non‑linear allocation patterns of IPv6 and transient mobile networks. IPGeoAI reframes the problem as a sequential modeling task. The core uses a Transformer Encoder to capture hierarchical dependencies in IP subnet structures. To resolve geographic ambiguity we employ a Zero‑Shot LLM feature extraction pipeline that offline converts noisy Autonomous System (AS) descriptions into structured domain tags such as “University” vs. “ISP” or “Global” vs. “Local”. These semantic signals are fused into the model via a Multi‑Head Cross‑Attention module, bridging numerical topology with real‑world identity. A hierarchical inference strategy first extracts coarse country signals and then refines them to city granularity, achieving 100% traffic coverage. On a proprietary dataset covering 200,000 cities, IPGeoAI outperforms a leading vendor by 6% in city‑level accuracy and yields a statistically significant +0.35% lift in a primary downstream metric during large‑scale online tests.
Review