Phones in dense 5G networks switch between cells so often that the switching itself becomes the bottleneck. Each switch, or handover, costs a round of control signaling and a short interruption. One proposed fix is handover skipping: stay on the current cell and deliberately pass over a small cell you would only occupy for a moment. Published skipping rules decide which cells to skip from the user's predicted path and the size of each cell, and they assume the network sees this information cleanly everywhere. It does not. The measurement reports phones send back go missing most often where the signal is weakest, at cell edges, in low-power small cells, and on sparse suburban corridors. Those are exactly the places where a wrong skip decision is most dangerous.
Epidemiology has a name for this pattern: ascertainment bias. A disease registry's silence often reflects missing diagnostic infrastructure, not missing disease. We bring that lens to mobility management. Our claim is that skipping rules evaluated on complete telemetry overstate their benefit, and that the overstatement is concentrated in the poorly reported regions. The cause is that a failed skip is not a small loss. The phone loses its link, falls back to blind channel scanning, and suffers an interruption that can be an order of magnitude longer than the handover it tried to avoid. We propose a skip rule that treats missing reports as informative rather than random, weighs the skip against this worst-case recovery latency, and is judged region by region rather than on network-wide averages.
Core Claim
A skip decision should compare two costs. One is the certain cost of one ordinary handover. The other is the expected cost of skipping, which includes a small probability of catastrophic recovery. Both the dwell-time estimate and the failure probability should be computed with a model of which reports the network failed to receive.
Let $D_h$ be the interruption of a normal prepared handover. Let $D_c$ be the interruption after radio link failure and cold rescanning. Let $\Delta R$ be the rate lost by staying on the weaker cell over the estimated dwell time $\hat{T}$. Let $p_{\text{fail}}$ be the probability that the serving link breaks during the skip. The rule skips when
$$ p_{\text{fail}}\, D_c + \Delta R\,\hat{T} \;<\; D_h . $$
Here $\Delta R\,\hat{T}$ is expressed in equivalent interruption time. Existing location-aware skipping effectively sets $p_{\text{fail}}$ to zero and trusts $\hat{T}$. That is harmless where telemetry is dense. It is wrong where telemetry is thin, because thin telemetry both biases $\hat{T}$ and hides the signal decay that predicts $p_{\text{fail}}$. The ratio $D_c/D_h$ is large: reactive scanning is over 90% of legacy handover delay (Reducing Handover Latency in Mobile Wireless Networks). So even modest errors in $p_{\text{fail}}$ flip the decision.
The mechanism uses one observable fact. For periodic and scheduled measurement reports, the network knows a report was expected, so a missing report can be recorded as a missing report. We model the chance that a report arrives, $\pi(r, u)$, as a function of the last observed signal strength $r$ and the uplink condition $u$. We then estimate dwell time and link-failure risk with a likelihood that includes this arrival model. This is the wireless analogue of correcting incidence estimates with a model of registry coverage. It is deliberately a small statistical model, not a deep network.
Relation to Prior Work
Three stochastic-geometry papers supply the skipping schemes we build on and the simulation methodology we reuse. Handover Management in Dense Cellular Networks: A Stochastic Geometry Approach shows that skipping every second base station halves handover delay cost, but it treats handover time as a fixed constant. Velocity-Aware Handover Management in Two-Tier Cellular Networks defines velocity-dependent schemes, including femto skipping and ignoring the femto tier entirely. It reports up to 77% throughput gains at 80–200 km/h, with delay modeled as a lumped parameter. Handover Management in 5G and Beyond: A Topology Aware Skipping Approach introduces location-aware, cell-size-aware, and hybrid skipping, with up to 47% throughput gain in two-tier networks. It explicitly flags inaccurate trajectory estimation as a limitation. All three assume the information behind a skip decision is complete, and none models the latency of a skip that fails. We keep their skip triggers and their interference-cancellation and coordinated-multipoint assumptions. We change what the triggers are fed and what a failure costs.
The handover-latency literature constrains the cost model. Proactive neighbor reports cut handover delay by up to 92% (Reducing Handover Latency in Mobile Wireless Networks). Motion-sensor-assisted pre-scanning reaches 23.45 ms (Reducing Handover Latency in Mobile Wireless Networks). Voice needs interruptions under about 50 ms (Reducing Handover Latency in Mobile Wireless Networks). These results give us the gap between $D_h$ for a prepared handover and $D_c$ for cold recovery, and a concrete threshold for tail-latency violations.
The geographic reporting bias work motivates the lens but supplies no method. Global temporal patterns of pancreatic cancer and association with socioeconomic development notes that high-quality registries cover about a third of the world's population. Temporal trends in cross-country inequalities of early-onset pancreatic cancer shows apparent burden tracking development level in ways partly attributable to under-diagnosis. The lesson we take is procedural: missingness correlates with the quantity of interest, and aggregate numbers hide stratum-level failures. We do not claim any quantitative transfer from health data.
Two learned-graph methods are neighbors rather than foundations. Neighbor Auto-Grouping Graph Neural Networks for Handover Parameter Configuration in Cellular Network tunes handover thresholds from full-coverage metropolitan telemetry. Variational-LSTM Autoencoder to forecast the spread of coronavirus across the globe learns latent spatial graphs under uneven national reporting. Neither models an explicit arrival process for the data. We deliberately start with an explicit, auditable arrival model instead of a learned latent graph, because a skip decision needs a calibrated failure probability more than a better throughput forecast.
Research Program
The first component is a censored-telemetry simulator. We extend the Poisson point process single-tier and two-tier cellular models used by the skipping papers in three ways. First, measurement reports are dropped with probability $1-\pi(r,u)$. The dropout is sharpest near the receiver sensitivity floor, and density-dependent inhomogeneity makes some regions report sparsely. Second, handover interruption becomes two-state: $D_h$ when the target was prepared through neighbor information, and $D_c$ after link failure with reactive scanning. Third, we record per-region outcomes. The degree of censoring is a controlled knob, set from missing at random, through missing depending on signal strength, to missing depending on region. This lets us separate the effect of how much is missing from the effect of what drives the missingness.
The second component is the estimator and rule. We fit $\pi$ from the expected-but-missing report indicators. We compute $\hat{T}$ and $p_{\text{fail}}$ by integrating over plausible signal paths consistent with both received and missing reports, and we apply the inequality above. $D_h$ and $D_c$ are measured inside the simulator, not assumed.
The third component is a real-data check, where access allows. The metropolitan datasets used by the TAG-GCN paper contain handover logs from about 1,459 cells. We would request them to test one thing only: whether report arrival there actually depends on last observed signal and region. If it does not, the premise weakens and we say so. If the data cannot be obtained, the work remains simulation-only, and that limitation is stated.
Baselines
The naive reference is best-connected handover with neighbor-report preparation: always switch to the strongest cell, with no skipping. It sets the handover rate and interruption profile that skipping is meant to improve. It is run in the same simulator, on the same topologies and trajectories, at the same censoring levels.
The strongest existing method is hybrid and location-aware skipping from Handover Management in 5G and Beyond: A Topology Aware Skipping Approach. Its reported numbers under complete information are up to 47% average throughput gain in two-tier networks and 8% at 100 km/h single-tier. We run it in two versions. The oracle version receives full telemetry, and we first check that our simulator approximately reproduces the paper's gains. The realistic version receives the censored reports, with the last value carried forward. We found no reported result on Lacuna for any skipping scheme under censored or missing telemetry, so the censored numbers are ours to establish.
The velocity-aware schemes of Velocity-Aware Handover Management in Two-Tier Cellular Networks enter as additional baselines, with reported gains of up to 77%. So does fixed alternate skipping from Handover Management in Dense Cellular Networks: A Stochastic Geometry Approach.
The most dangerous baseline is our own construction: location-aware skipping with a safety margin. Its dwell-time threshold is raised by a margin that depends on report density, and it is tuned by grid search per density stratum. It gets the same tuning budget, measured in simulated trajectories, that we spend fitting $\pi$.
Ablations isolate the two ingredients. One keeps the censoring-aware estimate but drops the asymmetric recovery cost. The other keeps the cost but uses a missing-at-random imputation.
Beating the field means three things at matched or higher mean throughput. First, a lower rate of interruptions above 50 ms and fewer radio link failures than the censored location-aware scheme. Second, the same advantage over the tuned-margin baseline in the worst-reported stratum. Third, no loss in well-reported strata.
Why not just…?
Why not just rely on neighbor reports and never skip? Neighbor reports make each handover cheap, up to 92% less delay (Reducing Handover Latency in Mobile Wireless Networks), but they do not reduce how many handovers happen. The skipping papers show throughput gains from fewer handovers starting around 30 km/h. Our best-connected baseline already includes neighbor-report preparation, so this alternative is measured directly.
Why not just add a conservative margin where reports are sparse? This is the real threat. If a per-stratum tuned margin matches our rule, then the "censoring model" is an expensive way to pick a threshold. The tuned-margin baseline exists to settle this question. We expect it to lose for a specific reason: sparseness of reports is not the same as risk. A sparse region with strong but uplink-limited signal should still skip, while a dense region with a cell dying at its edge should not. A fixed margin cannot tell these apart. The experiment varies these two cases independently. If the margin wins, we report that sparse-aware tuning suffices, and the contribution reduces to the censored benchmark and the per-region evaluation.
Why not just impute missing reports and run existing skipping? Missing-at-random imputation fills the gaps with typical values. Our premise is that the gaps occur exactly where values are atypically poor, which is the lesson of ascertainment bias in Temporal trends in cross-country inequalities of early-onset pancreatic cancer. The imputation ablation tests this directly. In the random-missingness setting it should tie our method. In the signal-dependent setting it should overestimate dwell time and cause failed skips. If it ties everywhere, our premise is wrong.
Why not just ignore the small-cell tier for fast users? Ignoring femtocells, as in Velocity-Aware Handover Management in Two-Tier Cellular Networks, sidesteps small-cell telemetry altogether, but it gives up small-cell capacity. That paper itself reports femto skipping beating legacy handover by more than 10% at 100 km/h. We include this scheme and report macro-tier load alongside throughput, so its robustness is not mistaken for a free win.
Decisive Checks
The first check is reproduction: oracle location-aware skipping should approximately recover its reported two-tier gain in our simulator. The second check is the premise itself. As censoring moves from random to signal-dependent at a fixed fraction of missing reports, the realistic location-aware scheme should lose throughput and gain long interruptions, with the loss concentrated in the worst-reported stratum. If it degrades equally under random missingness, the reporting-bias framing adds nothing.
The third and central check compares against the tuned-margin baseline in the regime where report sparseness and link risk are decorrelated. The fourth reports every result per stratum, alongside network averages, as epidemiology does across development levels. The point is to see whether aggregate gains are being purchased with outages where the network sees least. We run velocities from 30 to 200 km/h. Metrics are throughput, handovers per kilometer, radio link failure rate, 95th-percentile interruption, and the fraction of interruptions above 50 ms.
References
- A Pre-Scanning Fast Handoff Scheme for VoIP in WLANs
- Handover management in dense cellular networks: A stochastic geometry approach
- Velocity-Aware Handover Management in Two-Tier Cellular Networks
- Handover Management in 5G and Beyond: A Topology Aware Skipping Approach
- An 802.11k Compliant Framework for Cooperative Handoff in Wireless Networks
- Secured Fast Handoff in 802.11-Based Wireless Mesh Networks for Pervasive Internet Access
- Global temporal patterns of pancreatic cancer and association with socioeconomic development
- Temporal trends in cross-country inequalities of early-onset pancreatic cancer: a comprehensive analysis for the global burden of disease study 2021
- Neighbor Auto-Grouping Graph Neural Networks for Handover Parameter Configuration in Cellular Network
- Variational-LSTM autoencoder to forecast the spread of coronavirus across the globe