PaperlayerView exact revision on arXiv

Source-linked AI summary

Handover Analysis for Vehicular Communication with Explainability on the Fly

Ali Fuat Sahin, Semiha Tedik Başaran, Tufan Kumbasar

arXiv:2608.14820v1eess.SPcs.AI

TL;DR

Vehicular handover detection needs fast, reliable decisions, but black-box machine-learning models provide limited insight into which KPIs drive them. This paper evaluates inherently interpretable fANOVA-based models against LSTM with post-hoc SHAP, finding competitive detection performance with substantially lower explanation latency.

  • Problem

    Machine-learning handover detectors capture complex KPI relationships but remain largely opaque, limiting insight into which KPIs drive decisions.

  • Method

    The study evaluates inherently interpretable fANOVA-based models on real-world operator data against an LSTM model enhanced with post-hoc explanations.

  • Results

    fANOVA-based models achieve competitive detection performance with orders-of-magnitude lower explanation latency than LSTM with post-hoc SHAP.

  • Takeaways & Limitations

    Inherently interpretable models offer an efficient, transparent approach for latency-sensitive vehicular handover detection and reveal physically meaningful KPI relationships.

  • Takeaways & Limitations

    The DoNext dataset provides only serving-cell metrics, constraining analysis when simultaneous serving- and neighboring-cell information is required.

Abstract

from arXiv · show

Handover (HO) management in vehicular networks requires fast and reliable decision-making under highly dynamic conditions. While machine learning (ML) approaches can improve HO detection by capturing complex relationships among various key performance indicators (KPIs), their black-box nature limits interpretability and operator trust. To address this, this paper investigates HO detection from an explainability-on-the-fly perspective using inherently interpretable models based on the functional analysis of variance (fANOVA) framework. The proposed models are evaluated using two real-world operator datasets and compared against a Long Short-Term Memory baseline augmented with post-hoc SHAP explanations. Unlike post-hoc approaches, the proposed framework enables immediate interpretation of model decisions without incurring additional computational overhead. This capability is particularly critical for latency-sensitive vehicular networks. The results show that fANOVA-based models achieve competitive detection performance while providing significantly reduced explanation latency compared to conventional post-hoc methods. Furthermore, feature ranking and visualization analyses reveal physically meaningful relationships between KPIs and HO occurrences that align with standardized HO mechanisms. These results demonstrate that inherently interpretable models provide an efficient and transparent solution for HO detection in next-generation vehicular networks.

I. INTRODUCTION · II. SYSTEM MODEL

The paper addresses explainable handover detection for highly dynamic vehicular 5G NR networks, where fast decisions and operator trust are both essential. It combines inherently interpretable fANOVA-based models with a 5G NR mobility model grounded in standardized measurement events and KPI thresholds.

  • I. INTRODUCTION: Next-generation wireless applications such as connected and autonomous vehicles require ultra-reliable, low-latency communication in highly dynamic environments.Network densification, heterogeneous deployments, and dynamic conditions further complicate network design.
  • I. INTRODUCTION: Handover mechanisms maintain seamless cellular connectivity across coverage areas, motivating stochastic, game-theoretic, fuzzy-logic, and learning-based optimization approaches.These approaches address mobility challenges under stringent performance requirements.
  • I. INTRODUCTION: XAI improves transparency by quantifying feature contributions and revealing input-output relationships, thereby supporting interpretability, trust, and accountability.Wireless-communication applications include KPI analysis and anomaly detection.
  • I. INTRODUCTION: The study evaluates inherently interpretable fANOVA-based models for HO detection in vehicular 5G NR networks against a conventional ML model with post-hoc explanations using real-world operator measurements.The evaluation examines predictive performance, interpretability, and explanation latency while capturing realistic HO behavior.
  • I. INTRODUCTION: Intrinsically interpretable models achieve competitive predictive performance while significantly reducing explanation latency compared to post-hoc methods.They also identify influential KPIs and reveal physically meaningful relationships with HO occurrence.
  • II. SYSTEM MODEL: The system model contains a serving 5G NR gNB, neighboring gNB and LTE eNB, multiple UEs, and a 5G Core Network coordinating mobility and connectivity functions.The AMF manages HO and paging procedures, while the UPF manages user connectivity.
  • II. SYSTEM MODEL: 3GPP-defined A-series events govern same-RAT mobility, while B-series events govern inter-RAT mobility; practical examples include A3, A5, and B2.A3 typically supports intra-frequency handovers, A5 inter-frequency handovers, and B2 gNB-to-eNB inter-RAT handovers.
  • II. SYSTEM MODEL: HO trigger conditions compare serving and neighboring-cell metrics against predefined thresholds using MSC, MNC, and MNiRC.Representative KPIs include RSRP, RSSI, and RSRQ, expressed in dB or dBm, with thresholds T1 and T2 in corresponding units.

III. EXPLAINABLE MODELS & HANDOVER DATASETS

The section introduces inherently interpretable models for analyzing handover behavior and an LSTM classifier as a predictive baseline. SHAP provides post-hoc explanations for the LSTM’s handover predictions.

  • Explainable models: Inherently interpretable models are used to investigate the underlying behavior of the handover process.The section also considers handover datasets for this analysis.
  • Explainable models: The LSTM classifier models temporally characterized handover events using P KPIs and sequence length T.Its recurrence updates hidden and cell states across t = 1, . . . , T.
  • Explainable models: SHAP is adopted as the representative post-hoc method because it is established in wireless communications, model-agnostic, and capable of feature ranking.The LSTM output layer maps the final hidden state to predicted handover probability g(X).

A. Explainable Models

The study’s inherently explainable models use the fANOVA framework to represent responses through additive feature effects and interactions. It implements GAMs, EBMs, and GAMI-Net as representative approaches.

  • fANOVA framework: All built-in explainable models considered in the study fall under the fANOVA framework, which decomposes responses into additive terms of increasing order.The framework represents feature contributions through terms of increasing order.
  • fANOVA framework: First-order terms represent main effects of individual features, while second-order terms represent interactions between feature pairs.Higher-order interactions can also be included, trading off model complexity against interpretability.
  • Implemented models: The implemented representative approaches are Generalized Additive Models (GAMs), Explainable Boosting Machines (EBMs), and GAMI-Net.These three approaches are introduced as representative fANOVA-based models.

1) GAM:

GAMs provide a simple, interpretable fANOVA model that captures nonlinear main effects through flexible smooth terms. However, they cannot represent higher-order interactions, which may reduce performance as data complexity increases.

  • GAM: GAMs use flexible smooth terms to capture nonlinearities while retaining interpretability.They are described as one of the simplest fANOVA models.
  • GAM: GAMs model only main effects and cannot capture higher-order terms, potentially limiting performance as data complexity grows.This limitation motivates the use of more advanced models.

2) Explainable Boosting Machines:

Explainable Boosting Machines (EBMs) extend the GA2M framework into specialized gradient-boosting models with smooth, regularized shape functions and transparent visualizations. GAMI-Net further encodes fANOVA directly into a neural architecture through additive main-effect and interaction subnetworks.

  • Explainable Boosting Machines: EBMs build on GA2Ms as specialized gradient-boosting models that learn smooth, regularized spline-like shape functions.GA2Ms augment GAMs with pairwise interaction terms representing nonlinear main effects and salient interactions.
  • Explainable Boosting Machines: EBMs provide one-dimensional main-effect plots and two-dimensional heatmaps for pairwise interactions while achieving competitive accuracy.
  • GAMI-Net: GAMI-Net encodes the fANOVA expression directly into a neural architecture using dedicated subnetworks for main effects and selected interactions combined additively.It incorporates three interpretability constraints to enhance predictive performance while maintaining transparency.

3) GAMI-Net: · B. Handover Datasets

The GAMI-Net section retains active main effects and interaction pairs, while the handover analysis uses two public real-world datasets because radio-link metrics for both the serving and neighboring cells are required. A windowing approach converts these datasets into samples suitable for handover analysis.

  • 3) GAMI-Net:: GAMI-Net retains only active main effects S1 and interaction pairs S2.The notation μ, f_i(·), and f_ij(·) follows equation (5).
  • 3) GAMI-Net:: Table II presents the KPIs used for the handover analysis.
  • B. Handover Datasets: The handover analysis requires radio-link metrics for the serving cell and neighboring cell simultaneously.This requirement is shown in (1a, 1b, 1c).
  • B. Handover Datasets: Many existing datasets emphasize network-level metrics rather than radio-link performance.
  • B. Handover Datasets: Most handover studies rely on simulation data instead of real-world measurements.
  • B. Handover Datasets: Two publicly available datasets are employed to address limitations in metric coverage and data realism.The datasets are cited as and.
  • B. Handover Datasets: A windowing approach is applied to both datasets to obtain samples suitable for handover analysis.

1) Beyond Throughput Dataset: · 2) DoNext Dataset: · IV. COMPARATIVE PERFORMANCE ANALYSIS

The study uses mobility-focused samples from two real-world operator datasets to capture handover events and evaluates classification and interpretability performance. The DoNext subset contains 59,472 handover events across 152,775 samples, while its SC-only metrics limit the analysis.

  • 1) Beyond Throughput Dataset:: The Beyond Throughput dataset comes from a major Irish operator and was designed with comprehensive KPIs, mobility scenarios, and diverse application use cases.Only driving-scenario samples are used in this study to observe handover events.
  • 1) Beyond Throughput Dataset:: Beyond Throughput samples were restricted to the driving scenario for handover-event analysis.
  • 2) DoNext Dataset:: The DoNext dataset combines fleet and public-order vehicle campaigns on geofenced routes with long-duration, rail, and stationary measurements across Dortmund.These campaigns produced an open dataset spanning mobile, static, and rail scenarios.
  • 2) DoNext Dataset:: DoNext includes radio-link quality KPIs plus active data-rate and latency measurements, but this study uses only mobility-use-case data to capture handovers.
  • 2) DoNext Dataset:: 59,472 handover events across 152,775 samples yield a handover ratio of 38.93% in the DoNext dataset.
  • IV. COMPARATIVE PERFORMANCE ANALYSIS: The experiments evaluate both classification and interpretability performance.

A. Experiment Design

The experiment formulates handover detection as binary classification focused on identifying and explaining KPI conditions associated with handover decisions. It evaluates interpretable GAM, EBM, and GAMI-Net models against an LSTM baseline trained with specified optimization settings.

  • Task formulation: Handover detection is defined as binary classification of KPI conditions associated with handover decisions, rather than triggering or predicting future handovers.The task targets on-the-fly interpretability by identifying and explaining conditions linked to handover decisions.
  • Candidate models: Standard GAM, EBM, and GAMI-Net implementations are adopted using the configurations reported in.These models provide the interpretable candidates for the experiment.
  • Baseline model: An LSTM with 50 hidden units is trained using Adam, binary cross-entropy, and early stopping as the baseline model.Its predictions are assessed using SHA…, as stated in the supplied passage.

B. Handover Detection Analysis

Across two operator datasets, LSTM generally delivers the strongest predictive performance, while fANOVA-based models—especially EBM—provide lower explanation latency for on-the-fly handover detection. BT yields higher average AUC and F1-score than DoNext, suggesting benefits from NC measurements.

  • BT dataset: Over 10 independent runs, LSTM achieves the strongest BT classification performance, with test accuracy 0.914, AUC 0.938, precision 0.879, recall 0.862, and F1-score 0.870.EBM reaches training accuracy 0.946 but test accuracy 0.835, suggesting overfitting.
  • DoNext dataset: For DoNext, LSTM attains the highest test accuracy 0.804 and precision 0.792, while GAM leads AUC at 0.867 and GAMI-Net leads recall 0.698 and F1-score 0.728.EBM achieves training accuracy 0.866 but test accuracy 0.781; the relatively small metric differences indicate more balanced performance.
  • Accuracy-latency trade-off: EBM consistently achieves the lowest explanation latency among the fANOVA-based models, although this efficiency accompanies reduced predictive performance, particularly for BT.The accuracy-latency comparison uses AUC and F1-score bar plots and explanation latency in milliseconds per sample.
  • Dataset characteristics: BT records higher average AUC and F1-scores, 0.9198 and 0.7847, than DoNext, 0.8644 and 0.7180, respectively.The results suggest that incorporating NC measurements benefits reliable handover detection.

C. Explainability Analysis

The explainability analysis shows that HO decisions are mainly driven by radio-related KPIs and that inherently interpretable models expose physically meaningful relationships among distance, signal quality, and HO probability. GAMI-Net visualizations support the expected cause–effect behavior between UE movement, RSRQ, and HO occurrence.

  • Feature-based ranking: Signal strength-related KPIs, including RSRP, RSRQ, CQI, and their variations, consistently dominate the models’ HO decision rankings.EBM and GAMI-Net primarily select radio-related KPIs, while GAM and LSTM+SHAP sometimes include contextual features such as Cell Index and TrackingAreaCode.
  • Relationship-based visualization: The analysis derives a direct analytical association between Distance and RSRQ using received-power relationships and the Friis transmission equation.The formulation omits small-scale fading, frequency-dependent terms, and antenna gains for analytical simplicity.
  • Relationship-based visualization: As the UE moves away from the serving cell, RSRQSC degrades while proximity to the neighboring cell improves RSRQNC, supporting HO triggering.This relationship is interpreted geometrically for a UE moving between a serving cell and a neighboring cell.
  • Relationship-based visualization: GAMI-Net shows Distance increasing monotonically with P(HO), while RSRQ has a negative relationship with HO probability.The Distance relationship is consistent with signal attenuation models in which received power decays with squared distance; lower RSRQ increases HO likelihood.

V. CONCLUSION

The paper compared inherently interpretable fANOVA models with an LSTM baseline using post-hoc SHAP for HO detection in vehicular 5G NR networks. Although LSTM achieved strong detection performance, fANOVA models offered competitive performance with lower explainability overhead for latency-critical scenarios.

  • Conclusion: The study evaluated HO detection through an explainability-on-the-fly lens, comparing fANOVA models with an LSTM baseline using post-hoc SHAP explanations.The comparison targeted vehicular 5G NR networks.
  • Conclusion: The LSTM achieved strong detection performance, but post-hoc explainability introduced significant computational overhead that limited suitability for latency-critical HO scenarios.The passage identifies computational overhead as the principal limitation of the LSTM approach.
  • Conclusion: fANOVA-based models provided competitive detection performance with orders-of-magnitude lower explainability overhead than the LSTM baseline.This supports their use for explainability during latency-critical handover decisions.
Loading 2608.14820v1…