Predictive Quality of Service (PQoS)

• Predictive Quality of Service (PQoS) is the capability to estimate future network performance using current measurements and contextual signals to drive proactive adaptation.
• PQoS employs diverse methods including supervised learning, reinforcement learning, probabilistic modeling, and tensor-based approaches to forecast metrics such as throughput, latency, and violation probability.
• Applications of PQoS span vehicular networks, tele-operated driving, and cloud services, enabling systems to preemptively adjust operations and enhance quality assurance.

Searching arXiv for recent and foundational papers on Predictive Quality of Service to ground the article in current research. Predictive Quality of Service (PQoS) denotes the capability to estimate future communication or service performance sufficiently ahead of use so that networks, applications, or cyber-physical systems can adapt proactively rather than react after degradation has already occurred. In the literature, PQoS appears in several technically distinct forms: short-horizon network analytics exposed to applications in 5G vehicular systems, route-aware throughput forecasting from prior vehicles’ radio measurements, probabilistic forecasting of service-level violations via stochastic model checking, and temporal completion of sparse QoS tensors in web-service ecosystems (Ain et al., 2024, Kousaridas et al., 2021, Cicotti et al., 2014). Across these settings, the common abstraction is a mapping from current observations, contextual signals, and historical traces to future QoS quantities such as throughput, latency, reliability, response time, or violation probability, with the prediction then driving scheduling, compression, service selection, routing, or safety actions.

1. Conceptual scope and formalization

PQoS in vehicular and cellular networking is defined as estimating future communication quality that a moving user will experience over a cellular network, with horizons from seconds to minutes depending on the use case (Ain et al., 2024, Kousaridas et al., 2021). In the vehicular LTE setting of route-based throughput prediction, the prediction target is downlink throughput at a future time or location, modeled as a supervised regression problem

$y'(t+\tau)=\hat{f}(x(t),\theta),$

where $x(t)$ is a feature vector built from current measurements and contextual signals, $\theta$ denotes model hyperparameters, and $y'(t+\tau)$ is the predicted future throughput (Ain et al., 2024). In tele-operated driving, the same conceptual structure is expressed as forecasting expected values of QoS profile parameters and their related variances over a prediction horizon $t_{\text{pred}}$, with the horizon tied to vehicle dynamics and safe maneuver requirements (Kousaridas et al., 2021).

A second formulation treats PQoS as a stochastic forecast of future service states. In probabilistic model-checking approaches, a KPI is discretized into admissible, critical, and inadmissible regions, and a parametric CTMC is updated from runtime data to compute quantities such as

$\mathcal{P}_{=?}[F_{\le T}\;\text{"violState"}],$

that is, the probability that a violation state will be reached within time $T$ (Cicotti et al., 2014). This formulation shifts PQoS from point prediction to quantified future risk.

A third formulation arises in web-service and cloud settings, where PQoS is prediction of missing or future user-specific QoS values in sparse user–service–time data. Dynamic QoS is represented as a tensor $Y \in \mathbb{R}^{|I|\times|J|\times|K|}$, and the task is to infer $\hat{Y}_{ijk}$ for unobserved user–service–time triples (Wang et al., 2024). In more recent temporal latent-feature models, the same setting is represented as a sequence of user–service matrices $Y_{(t)}$, with temporal latent user features $x(t)$0 and time-invariant service features $x(t)$1, yielding predictions of the form

$x(t)$2

(Yuan et al., 22 Jun 2026).

A broader systems-level view defines PQoS as prediction plus exposure plus application adaptation. Under that interpretation, PQoS is not merely a forecasting model but a control primitive: the network predicts future QoS, exposes it through standardized or internal interfaces, and the application or controller changes behavior accordingly (Kousaridas et al., 2021, Boban et al., 2021).

2. Architectural realizations

In 5G V2X, the canonical standardized realization of PQoS is based on the NEF–NWDAF architecture. A V2X Application Server declares QoS requirements, location or geographical area, a time window, and thresholds for QoS deviation; the NEF exposes analytics APIs; the NWDAF gathers measurements from OAM and computes QoS sustainability analytics; the resulting predictions are notified back to the application, which can then adapt (Kousaridas et al., 2021). The relevant analytics interface is “QoS sustainability analytics,” and the procedure supports both subscription and one-shot fetch modes through Nnef_AnalyticsExposure_Subscribe, Nnef_AnalyticsExposure_Fetch, and Nnef_AnalyticsExposure_Notify (Kousaridas et al., 2021). This architecture makes explicit that PQoS in mobile systems is as much about exposure semantics and integration as about prediction accuracy.

A complementary line places PQoS logic at the RAN. In teleoperated-driving scenarios, a RAN-level entity termed RAN-AI ingests multi-layer measurements and selects application countermeasures, such as LiDAR compression modes, through RL (Mason et al., 2022). In decentralized variants, similar agents can be pushed to vehicles, with centralized, distributed, and federated learning used as alternative coordination schemes (Bragato et al., 2023). This suggests a bifurcation of PQoS architecture into core-centric analytics exposure and low-latency RAN- or endpoint-centric control.

Outside mobile networking, PQoS architectures can be built around monitoring, parameter estimation, formal modeling, and verification. The probabilistic model-checking framework for QoS prediction uses CEP to aggregate events, estimate model parameters such as $x(t)$3 and $x(t)$4, instantiate a PRISM model, evaluate CSL properties, and convert them into Quality Constraints over predictive indicators (Cicotti et al., 2014). This pipeline is architecturally different from ML-based approaches but still satisfies the same predictive loop.

In service ecosystems, PQoS architectures are usually data-centric rather than interface-centric. Historical QoS logs are collected, organized into matrices or tensors, factorized or forecast, and then used for service selection or orchestration (Wang et al., 2024, Yuan et al., 22 Jun 2026). A plausible implication is that PQoS in these environments is often embedded into ranking or recommendation engines rather than exposed as a standalone network analytic.

3. Prediction targets, horizons, and metrics

The prediction targets used in PQoS studies vary substantially by domain. In route-aware vehicular prediction, the primary target is downlink throughput, specifically maximum achievable throughput at a future time or location, with PHY and cell-level measurements such as SNR, RSRP, RSRQ, RSSI, cell load, and number of connected devices serving as features (Ain et al., 2024). In tele-operated driving feasibility analyses, the evaluation focuses on UL throughput, while QoS profiles also include latency and service-mode-specific rate requirements such as full video at $x(t)$5, limited video at $x(t)$6, reduced video at $x(t)$7, slim uplink at $x(t)$8, and control downlink at about $x(t)$9 with latency constraints of $\theta$0 or $\theta$1 depending on direct or indirect control (Kousaridas et al., 2021).

Prediction horizons are correspondingly heterogeneous. Tele-operated driving in the 5G QoS sustainability framework is explicitly short-term: $\theta$2 depends on vehicle speed and safe stopping requirements, ranging from about $\theta$3 at $\theta$4 to about $\theta$5 at $\theta$6 (Kousaridas et al., 2021). In contrast, vehicular route-based throughput prediction with prior users operates on minute-scale horizons. The effective horizon can be defined through inter-vehicle distance and speed,

$\theta$7

with reported average lags around $\theta$8 minutes and $\theta$9 minutes for different lead vehicles (Ain et al., 2024). This demonstrates that PQoS is not tied to a single timescale; rather, the horizon is application-dependent and constrained by the predictability of the underlying environment.

Evaluation metrics also differ. For route-based throughput prediction, the main measure is Mean Relative Percentage Error, defined theoretically as

$y'(t+\tau)$0

with a practical denominator modification using $y'(t+\tau)$1 for low-throughput cases (Ain et al., 2024). Tele-operated-driving feasibility studies use MAE, standard deviation of absolute error, and MAPE for UL throughput prediction (Kousaridas et al., 2021). Tensor-based web-service prediction studies typically use RMSE and MAE on held-out entries (Wang et al., 2024). Formal verification approaches instead quantify future violation probability directly, using temporal-logic expressions and threshold comparisons rather than regression losses (Cicotti et al., 2014).

4. Methodological families

One major family is supervised ML for future QoS regression. In vehicular LTE prediction with prior users, XGBoost is used as the regressor because earlier work on the same dataset had found tree boosting to outperform neural networks and random forests for data-rate prediction (Ain et al., 2024). The feature sets range from current self-throughput alone to self PHY, self PHY plus cell indicators, next-vehicle PHY, and next-vehicle PHY plus self-cell information, all trained separately for different inter-vehicle horizons (Ain et al., 2024). The study explicitly reports that feature importance is dominated by next-vehicle PHY, while self-vehicle PHY contributes little, and that adding all self and next features together worsens performance due to overfitting on the small dataset (Ain et al., 2024).

A second family uses explicit forecasting of input features combined with a learned QoS model. In the 5G connected and automated driving framework, a Random Forest regressor maps position, serving-BS distance, multi-cell load, demand, and interference proxies to UL throughput, while ARIMA is used online to forecast future load-related features over the prediction horizon (Kousaridas et al., 2021). This decomposes PQoS into prediction of state variables and prediction of QoS conditional on those state variables.

A third family embeds prediction into reinforcement learning. In RAN-level teleoperated-driving studies, PQoS is realized as an RL loop in which the learned Q-function or policy implicitly predicts the future consequences of actions on latency, PRR, and application quality. The reward can couple QoS and QoE, for example through a weighted combination of delay margin and Chamfer Distance when latency and PRR constraints are satisfied, and zero otherwise (Mason et al., 2022). Later work extends this to centralized, distributed, and federated schemes, using the learned value function as the predictive mechanism without an explicit forecast model (Bragato et al., 2023). More recent teleoperated-driving formulations compare MAB, SARSA, Q-Learning, DSARSA, and DDQN under federated learning and report that linear Q-Learning is the best trade-off in average reward, convergence, and computational cost for their TD scenario (Bragato et al., 2024). A still more recent extension introduces two integrated RL agents for joint compression and scheduling, together with a meta-learning agent that chooses between centralized and decentralized deployment strategies depending on network conditions and application requirements (Avanzi et al., 24 Mar 2026). This suggests a broadening of PQoS from forecasting-only to learned proactive control.

A fourth family is probabilistic formal modeling. The CTMC-based framework parameterizes KPI dynamics via increment and decrement rates estimated from runtime data, then uses PRISM to compute probabilities of future critical or violation states (Cicotti et al., 2014). This approach differs from ML in that the predictive object is not a scalar estimate but a model-checked probability, enabling predictive SLAs and pre-alerts such as rejecting configurations when the probability of violation within a future bound exceeds a threshold (Cicotti et al., 2014).

A fifth family covers tensor and latent-factor methods for temporal QoS data. The Extended Canonical Polyadic-based Tensor Network introduces a relation dimension $y'(t+\tau)$2 between user and service latent spaces so that

$y'(t+\tau)$3

and is trained by a nonnegative multiplicative update algorithm on incomplete tensors (Wang et al., 2024). More recent temporal latent-feature work combines an Extended Kalman Filter over temporal user features with ALS-based time-invariant service features, explicitly targeting non-stationary fluctuations (Yuan et al., 22 Jun 2026). These methods instantiate PQoS as dynamic tensor completion rather than explicit network control.

A sixth family combines robust learning, graph structure, and temporal modeling. The Hypergraph Convoluted Transformer Network frames temporal QoS as prediction over a sparse user–service–time tensor, uses hypergraph convolution to capture high-order collaborative structure, includes a greysheep-detection module, and trains with Cauchy loss after optional Isolation Forest outlier removal (Kumar et al., 2024). This indicates that anomaly resilience has become a distinct subtheme within temporal PQoS.

5. Prior-user information, spatial correlation, and multi-entity context

A prominent result in vehicular PQoS is that measurements from prior users can serve as a proxy for the future radio environment. In the LTE drive-test study, Pearson cross-correlation of PHY features between trailing and leading vehicles peaks at about $y'(t+\tau)$4 minutes for the self–vehicle-3 pair and about $y'(t+\tau)$5 minutes for the self–vehicle-1 pair, matching the average temporal gaps between vehicles (Ain et al., 2024). This establishes that leading-vehicle measurements can approximate the future measurements of the trailing vehicle along the same route. The best-performing model, “Next PHY {data} Cell,” achieves the lowest MRPE and is reported as roughly $y'(t+\tau)$6 lower than the baseline current-QoS-only model at both $y'(t+\tau)$7- and $y'(t+\tau)$8-minute horizons (Ain et al., 2024).

Urban V2X work extends this principle from highway LTE to urban 5G/V2X environments using the Berlin V2X Cellular Dataset. There, downlink throughput prediction for an ego vehicle improves when lead-vehicle data and historical trends are added, and the improvement is described as model-agnostic across XGBoost, CNN, and LSTM (Partani et al., 23 Apr 2025). The best configuration, spatially aligned ego-plus-lead data with XGBoost, reduces normalized MAE from $y'(t+\tau)$9 to $t_{\text{pred}}$0, SMAPE from $t_{\text{pred}}$1 to $t_{\text{pred}}$2, and RMSE from $t_{\text{pred}}$3 to $t_{\text{pred}}$4, corresponding to improvements of $t_{\text{pred}}$5, $t_{\text{pred}}$6, and $t_{\text{pred}}$7, respectively, over the ego-only baseline (Partani et al., 23 Apr 2025). This strengthens the general claim that PQoS accuracy benefits from exploiting spatial-temporal correlation across users traversing similar contexts.

A related but distinct notion of multi-entity context appears in concurrent real-time communication flows. The Packet-to-Prediction framework processes multiple concurrent RTP flows jointly from raw packets and predicts four per-flow QoS metrics—bitrate, average jitter, FPS, and binary loss condition—over the next $t_{\text{pred}}$8, using a length-free Transformer with flow-wise cross attention and neighbourhood attention (Song et al., 2024). This addresses a limitation of per-flow predictors by capturing inter-flow dependence in shared network conditions. A plausible implication is that, in multi-flow scenarios, the “prior-user” idea generalizes to “peer-flow” context, where the state of one flow contains predictive information about others because they share bottlenecks.

6. Application domains and operational use

The most developed mobile-network application is tele-operated driving. In 5G architecture studies, predicted QoS can trigger changes from direct to indirect control, reduce vehicle speed, adapt video configuration, or initiate a smooth stop before QoS degrades (Kousaridas et al., 2021). In the broader autonomous-systems perspective, PQoS increases application survival time by providing advance notice of network service failures, allowing mode adaptation before the application enters a failure state (Boban et al., 2021). RL-based teleoperated-driving frameworks concretize this by mapping predicted or implicitly anticipated QoS to LiDAR compression modes that trade latency against point-cloud quality (Mason et al., 2022, Bragato et al., 2023, Bragato et al., 2024, Avanzi et al., 24 Mar 2026).

Other vehicular uses include platooning, cooperative maneuvers, route planning, and smart navigation. Minute-scale route-aware throughput forecasts can support pre-buffering, proactive handovers, switching RATs, route changes, or safe-stop preparation when upcoming connectivity loss is expected (Ain et al., 2024). In high-density platooning, anticipated out-of-coverage or low-QoS zones can motivate changes in platoon size, spacing, or control mode (Boban et al., 2021).

In web-service ecosystems, PQoS primarily supports service selection and recommendation. Temporal tensor models predict response time or throughput for user–service–time triples so that a system can rank functionally equivalent services by expected QoS without exhaustive invocation (Wang et al., 2024). Probabilistic QoS-profile approaches go further by representing service QoS as a multivariate continuous distribution and expressing service requirements as probabilistic constraints over regions $t_{\text{pred}}$9, with satisfaction determined by whether

$\mathcal{P}_{=?}[F_{\le T}\;\text{"violState"}],$0

holds for the service’s QoS profile (Suñé et al., 2022). This suggests that one strand of PQoS is less about dynamic control and more about probabilistic admissibility and selection under uncertainty.

In cloud and service-oriented systems, temporal QoS prediction supports resource allocation, SLA management, and missing-value completion. EKF-based latent-feature methods are motivated by the need to handle fluctuations in non-stationary temporal QoS data so that systems can rationalize resource allocation and optimize service selection more accurately (Yuan et al., 22 Jun 2026). Hypergraph–Transformer methods similarly target adaptive service recommendation in the presence of sparsity, anomalies, and atypical users or services (Kumar et al., 2024).

A non-communication example appears in mobility-on-demand systems, where predictive positioning uses a multinomial forecast of future arrivals to minimize expected wait time, and a random-forest customer-rating model maps predicted service metrics to perceived QoS (Miller et al., 2016). This suggests that “Predictive QoS” is broader than packet networks: it can denote any architecture that predicts future service quality under candidate decisions and optimizes accordingly.

7. Limitations, controversies, and open directions

A recurrent limitation is dependence on context availability. Prior-user vehicular methods require recent measurements from vehicles that traversed the same route; sparse traffic, rare routes, or stale data weaken the approach (Ain et al., 2024). This suggests that a deployment-grade PQoS system may need a connectivity map or network data lake, along with freshness criteria and trajectory matching logic, to determine when prior-user information is reliable enough to use.

Another limitation is generalization across environments. The LTE prior-user study is explicitly a first step in highway environments, with the authors planning extension to rural and urban settings where multipath and non-stationarity are more complex (Ain et al., 2024). The urban Berlin V2X study partially addresses that gap, but remains tied to a specific dataset and operator scenario (Partani et al., 23 Apr 2025). Similar concerns arise in teleoperated-driving RL work, where policies are often learned in single-cell or constrained simulated settings and may not transfer directly across cities, load regimes, or radio configurations (Mason et al., 2022, Bragato et al., 2023).

Prediction horizon versus accuracy remains a central tension. The 5G ToD analysis notes that short-term horizons up to about $\mathcal{P}_{=?}[F_{\le T}\;\text{"violState"}],$1 seconds are feasible, with errors increasing after roughly $\mathcal{P}_{=?}[F_{\le T}\;\text{"violState"}],$2 seconds in simulation (Kousaridas et al., 2021). In route-based minute-scale prediction, accurate forecasts are enabled precisely because leading vehicles provide a spatially aligned probe of future conditions (Ain et al., 2024). This suggests that longer horizons are tractable when there is a physical or social mechanism—such as repeated routes or shared trajectories—that injects advance information into the current state.

Data reliability is another contentious point. Many temporal QoS methods historically assumed reliable logs, but more recent work emphasizes outliers, greysheep users/services, and atypical invocation patterns (Kumar et al., 2024). Outlier-resilient matrix and tensor factorization based on Cauchy loss showed that large residuals can dramatically distort QoS prediction if treated with standard $\mathcal{P}_{=?}[F_{\le T}\;\text{"violState"}],$3 objectives, motivating robust losses and anomaly-aware preprocessing (Ye et al., 2020). A plausible implication is that, as PQoS systems become integrated into operational decision loops, robustness to mislabeled, adversarial, or distribution-shifted data will be as important as raw prediction accuracy.

Standardization is still incomplete. While NWDAF and NEF provide a standardized substrate for analytics exposure, RAN-side predictive intelligence, distributed sidelink-aware PQoS, and practical fusion of CN and RAN analytics remain open areas (Kousaridas et al., 2021, Boban et al., 2021). Federated and decentralized RL have been proposed partly to address privacy, overhead, and scalability, with federated learning often offering a favorable trade-off between convergence time and privacy preservation in vehicular networks (Bragato et al., 2023, Bragato et al., 2024). This suggests that future PQoS systems may be hybrid: standardized analytics exposure in the core, local prediction/control at the RAN or endpoint, and federated coordination across agents.

Methodologically, the field is broadening toward integrated control. Joint RL agents for compression and scheduling in teleoperated driving reportedly outperform standalone compression-only or scheduling-only models, especially under constrained or congested conditions (Avanzi et al., 24 Mar 2026). This suggests that PQoS is moving from prediction of single KPIs toward coordinated optimization of multiple knobs and multiple metrics, including latency, reliability, and task quality.

Taken together, the literature indicates that PQoS is best understood not as a single algorithmic technique but as a research program centered on anticipatory service assurance. Whether implemented through XGBoost on route-aligned vehicular data, Random Forest plus ARIMA in NWDAF, RL in teleoperated driving, CTMC model checking for predictive SLAs, tensor networks for web-service completion, or hypergraph–Transformer architectures for anomaly-resilient temporal prediction, the defining feature is the same: future QoS is estimated early enough, and with enough structural context, to change decisions before the service degrades (Ain et al., 2024, Kousaridas et al., 2021, Cicotti et al., 2014, Wang et al., 2024, Yuan et al., 22 Jun 2026, Kumar et al., 2024).