Inductive Latent Context Persistence: Closing the Post-Handover Cold Start in 6G Radio Access Networks

Anubhab Banerjee, Daniyal Amir Awan

Keywords: AI-native RAN, handover prediction, mobility management, heterogeneous graph neural networks, recurrent state transfer, O-RAN, near-RT RIC, robust learning, variational compression

TL;DR: ILCP transfers a compressed per-UE recurrent state over the 3GPP Xn interface at handover, eliminating the post-handover cold start and reducing ping-pong handovers in a real 5G drive-test trace.

Abstract: In modern radio access networks (RANs) rulebased handover (HO) decisions (e.g., A3/A5) depend on user equipment (UE) measurements only, resulting in different HO decisons for UEs who are in the same location. To overcome it, graph neural networks (GNNs) based methods have been proposed to improve HO key performance indicators (KPIs) using more information than just the measurements. However, existing recurrent or graph-based methods discard the per- UE recurrent state at HO and reinitialize it at the target next-generation NodeB (gNB) – losing useful measurement and mobility history as part of the overall context, forcing the target-side model to rebuild the state from post-HO measurements only. We address this post-HO cold start problem with Inductive Latent Context Persistence (ILCP), a learned latent synchronization mechanism that compresses the source-side per- UE recurrent state, transports it over the standard 3GPP Xn interface (Xn) as a 128-byte payload, and adapts it to the target gNB state space at HO. We model the RAN as a dynamic heterogeneous graph, allowing the model to treat UE nodes, gNB nodes, measurement edges, and Xn neighbor edges separately. On the Vienna 4G/5G drive-test, ILCP eliminates ping-pong HOs in the test split with 0.0% compared with 6.5% for the otherwise identical but no-transfer baseline and 22.6% for a Transformer baseline. It also achieves a +5.1 pp average post-HO accuracy gain, with a peak gain of +13.3 pp, over the no-transfer baseline in the 50–250 ms post-HO window. On a single NVIDIA GTX 1080 (8 GB), ILCP runs endto- end at 7.7 ms p99 per handover decision. Under measurement perturbations including shadow fading, non-line-of-sight (NLOS) blockage, and synchronization signal block (SSB)-burst sparsity, robustly trained ILCP keeps handover failure (HOF) in the 10–13% and exposes sub optimality of relying only on measurements. In the same fixed-reference-label setting, the 3GPP A3/A5 rule which relies only on measurements increases from 1.1% HOF on the unperturbed trace to 57– 65% under perturbed measurements.

Submission Number: 16