arXiv CSopen access

ResLearn-XR: Residual Learning for Network Traffic and Quality-of-Experience-Aware Modeling in Extended Reality

1

ResLearn-XR: Residual Learning for Network Traffic and Quality-of-Experience-Aware Modeling in Extended Reality Yoga Suhas Kuruba Manjunath, Member, IEEE,

arXiv:2609.04493v1 [cs.AI] 3 Sep 2026

Jie Gao, Senior Member, IEEE, Lian Zhao, Fellow, IEEE

Abstract—We present ResLearn-XR, a residual learning framework for predicting eXtended Reality (XR) network traffic and estimating Quality-of-Experience (QoE) risk. ResLearn-XR adopts a two-stage temporal learning structure comprising a base sequence prediction model augmented with task-specific residual learning components to improve adaptability to bursty, non-stationary XR traffic dynamics. The residual learning stages operate in the value space for continuous XR traffic forecasting and in the logit space for probabilistic QoE risk estimation. For the QoE-risk branch, we introduce a Data Descriptor Algorithm (DDA), a causal feature-construction module that converts packet-level application-layer observables into frametiming-aware descriptors suitable for encrypted traffic analysis. We also construct an XR Traffic-QoE dataset that pairs continuous XR traffic traces with session-level user-reported QoE labels. ResLearn-XR reduces SMAPE by up to 17.84% across frame-count, frame-size, and inter-arrival-time prediction, while reducing QoE-risk estimation SMAPE by up to 87.8% over single-stage baselines. Index Terms—Extended Reality, Quality of Experience, Residual Learning, Transformer Networks, Network Traffic Prediction, Motion-to-Photon Latency, Cognitive Communications, Intelligent Network Management

I. I NTRODUCTION Extended Reality (XR), encompassing Virtual Reality (VR), Augmented Reality (AR), and Mixed Reality (MR), represents one of the most demanding use cases for cognitive next-generation communication networks [1]. XR applications generate synchronized video, audio, and control traffic while operating under a strict motion-to-photon (MTP) latency constraint of under 20 ms [2]. Violations of this constraint trigger cybersickness, a sensory mismatch causing motion discomfort and severe Quality-of-Experience (QoE) degradation [3], [4]. Supporting XR over wireless networks, therefore, requires communication systems capable of cognitive behaviour: continuously observing traffic dynamics, inferring imminent QoE Y. Manjunath is with the School of Information Technology, Carleton University, Ottawa, ON, K1S 5B6, Canada, and the Department of Electrical, Computer and Biomedical Engineering, Toronto Metropolitan University, Toronto, ON M5B 2K3, Canada. (e-mail: [email protected], [email protected]). J. Gao is with the School of Information Technology, Carleton University, Ottawa, ON, K1S 5B6, Canada. (e-mail: [email protected]). L. Zhao is with the Department of Electrical, Computer and Biomedical Engineering, Toronto Metropolitan University, Toronto, ON M5B 2K3, Canada. (Corresponding author e-mail: [email protected])

risk, and enabling proactive resource management before MTP latency thresholds are breached. This cognitive intelligence requirement motivates two tightly coupled learning tasks. First, XR traffic prediction must capture both long-range temporal trends and the short-term, bursty frame-level fluctuations that directly drive network queueing delay and MTP violations. Second, QoE risk estimation must map observed network behaviour to the likelihood of elevated cybersickness risk, enabling predictive, rather than reactive, network control. Conventional temporal predictors, including time-series forecasters [5], Long Short-Term Memory (LSTM) networks [6], Gated Recurrent Unit (GRU) networks [7], and Transformer-based architectures such as Informer and FEDformer [8], [9], are trained with sequence-level regression objectives that bias predictions toward conditional mean behaviour. Consequently, high-variance frame-level traffic bursts are underpredicted, precisely at the time-scales critical for meeting MTP requirements. A single-stage predictor is thus insufficient to jointly model long-range trends and burst dynamics, motivating explicit residual learning for burst-phase correction. On the QoE estimation side, the ITU-T identifies XR QoE as dependent on network performance, configuration-dependent timing targets (e.g., display refresh rate), and user-perceived experience [10], [11]. Yet existing deep learning [12] and digital twin [13] approaches rely on static or aggregated network indicators, lacking the frame-level timing-variability representations needed to capture transient QoE degradation. Furthermore, existing XR datasets [14]–[17] provide traffic traces without associated session-level user QoE labels, preventing supervised learning of the network traffic-to-QoE relationship. We introduce ResLearn-XR, a two-stage residual learning framework for joint XR traffic prediction and QoE risk estimation. The framework augments a base Transformer encoder with task-specific residual heads: a value-space head corrects burst-phase traffic prediction errors, while a logitspace head refines probabilistic QoE risk estimates. DDA transforms encrypted-traffic-compatible packet-level observables into frame-timing-aware causal descriptors aligned with ITU-T QoE causality requirements. We additionally release an XR Traffic-QoE dataset in which session-level QoE annotations are paired with the corresponding continuous traffic

2

traces across diverse XR applications and network conditions. ResLearn-XR achieves causal SMAPE improvements of up to 17.84% (approximately 6.7% on average) across frame-count, frame-size, and inter-arrival-time prediction, and up to 87.8% SMAPE improvement for QoE risk estimation. Compared with our prior ResLearn traffic-prediction framework [18], which focused on frame-level XR traffic forecasting, this paper extends the scope to joint XR traffic prediction and QoE risk estimation through logit-space residual correction and DDA-based encrypted-traffic-compatible feature construction. It further introduces the XR Traffic-QoE dataset with session-level user annotations and expands the evaluation to weakly supervised QoE risk estimation, cross-user testing, DDA ablation, calibration, latency, and proxy-based MTP-risk analysis. This work makes three main contributions, • ResLearn-XR: a two-stage Transformer framework with value-space and logit-space residual heads for XR traffic prediction and QoE risk estimation, enabling cognitive awareness of burst dynamics and perceptual risk; • DDA: a causal feature construction method that derives frame-timing-aware causal descriptors from applicationlayer observables (packet length, direction, inter-arrival time, timestamp), enabling QoE risk estimation under encrypted traffic; • XR Traffic-QoE Dataset: a publicly released dataset covering diverse XR applications and system configurations with session-level user-reported QoE labels associated with the corresponding traffic traces, to support weakly supervised QoE risk estimation. To the best of our knowledge, ResLearn-XR is among the first frameworks to jointly address XR traffic prediction and QoE risk estimation in a unified cognitive architecture. The implementations for XR traffic prediction1 and QoE risk estimation2 are publicly available, as is the XR Traffic-QoE dataset3 .

requirements, motivating explicit residual modeling of shorttimescale prediction errors.

II. R ELATED W ORK

ftraf : RD×T → Rm ,

A. XR Traffic Prediction XR traffic exhibits piecewise non-stationary behavior driven by scene dynamics, encoder rate control, and rendering synchronization [19], undermining statistical models such as Autoregressive Integrated Moving Average (ARIMA) and Hidden Markov Models (HMMs) that assume stationarity. Vaidya et al. propose an LSTM-based transfer learning model for VR cloud-gaming traffic forecasting [6]; however, recurrent architectures remain limited in capturing abrupt frame-level bursts. Transformer-based models (Temporal Fusion Transformer [5], Informer [8], FEDformer [9]) improve long-range forecasting but oversmooth frame-scale bursts, underrepresenting the rapid transitions intrinsic to XR traffic, as confirmed by Morı́n et al. [20] and Chiariotti et al. [21]. These burst phases are critical for meeting frame delivery deadlines and MTP latency 1 https://github.com/yoga-suhas-km/ResLearn 2 https://github.com/yoga-suhas-km/XR QoE Prediction 3 https://dx.doi.org/10.21227/na8a-9n63

B. XR QoE Risk Estimation Ruan and Xie [22] identify that XR-critical timing factors such as frame-rate variability and MTP latency are only indirectly reflected in aggregated QoS metrics. Reactive approaches, cooperative rendering [23] and adaptive edge streaming [24], reduce latency but lack temporal learning mechanisms for predictive QoE risk estimation. Kougioumtzidis et al. [12] and Jiadong et al. [13] propose LSTMbased and digital-twin-based QoE models, but neither explicitly links XR traffic evolution to frame-level delivery behavior, limiting their ability to capture how XR traffic instability translates into elevated QoE risk. C. XR Traffic-QoE Datasets Existing XR traffic datasets [14]–[17], [19] provide packetand frame-level measurements but lack associated sessionlevel subjective QoE labels. Studies demonstrating the sensitivity of XR experience to transient network disruptions [20], [25], [26] use proprietary data. No open dataset jointly captures diverse XR traffic dynamics and corresponding sessionlevel user QoE outcomes, limiting development of predictive XR frameworks. Table I contextualizes ResLearn-XR within the literature. III. P ROBLEM F ORMULATION An XR session produces a network packet sequence X = {xt | t = 1, . . . , T }, where each xt ∈ RD encodes packetlevel attributes (direction, length, inter-arrival time, timestamp). The objective is to learn two predictive functions: (i) XR traffic prediction and (ii) QoE risk estimation, detailed below. Traffic prediction: We aim to learn a predictive function (1)

that captures the temporal evolution of XR traffic and forecasts its future state. The function operates on an observed packetlevel traffic sequence X as defined above. (1) (m) Let yt = [yt , . . . , yt ] ∈ Rm , m ≤ D, denote a vector of m frame-level traffic metrics, namely, frame count, average frame size, and frame inter-arrival time, derived by aggregating the underlying packet-level observations over a temporal window ending at time t. The dimensionality m satisfies m < D because the traffic metrics are not independent raw features, but lower-dimensional aggregates computed from the D packet-level features. While these metrics can be computed from packet traces after observation, they are not available at prediction time for future windows, as their computation requires packet arrivals beyond time T . Given the observed packet-level traffic sequence X up to bT +1 = time T , the predictor outputs a next-step estimate y ftraf (X), which aims to forecast the corresponding groundtruth traffic metrics yT +1 . The traffic prediction task learns

3

TABLE I C OMPARATIVE S UMMARY OF P RIOR W ORKS AND DATASETS IN XR T RAFFIC P REDICTION AND Q O E R ISK E STIMATION Reference

XR Traffic Prediction

[19] [6] [5], [8], [9] [20], [21] [22] [23], [24] [12] [13] [19], [14]–[17] [25], [20] This Work

Yes Yes Yes Yes No No No No No No Yes

XR QoE Risk Estimation No No No No Yes Yes Yes Yes No No Yes

XR Traffic-QoE Data No No No No No No No No No No Yes

Assumptions/Limitations

Assumes stationarity Limited temporal granularity; underfits non-stationary bursts Over-smooths short bursts; unsuitable for XR volatility Lack perceptual QoE outcomes Ignores temporal and perceptual variation in QoE Frame-level only features Perceptual-layer only; Latency-based features only Reactive orchestration No associated session-level subjective QoE labels Proprietary or scenario-limited -

the function ftraf by minimizing a task-specific loss Ltraf (·) between predicted and ground-truth future traffic states. QoE Risk Estimation: We aim to estimate the risk of QoE violation, hereafter referred to as “QoE risk,” associated with an XR traffic sequence. This task is formulated as a probabilistic mapping fqoe : RD×T → [0, 1],

(2)

which maps an observed packet-level traffic sequence X ∈ RD×T to a scalar risk probability pb = fqoe (X), where pb ∈ [0, 1] represents the estimated likelihood that the traffic sequence is associated with elevated QoE risk. QoE risk is encoded as a binary variable y ∈ {0, 1}, with y = 1 and y = 0 denoting elevated and non-elevated risk, respectively. We assign each session-level subjective QoE annotation to all temporal windows extracted from the corresponding XR session, formulating QoE-risk estimation as a weakly supervised temporal learning problem over packetlevel traffic observations. The task minimizes Lqoe = E[D(y, fqoe (X))] ,

(3)

where D(·, ·) is binary cross-entropy. The XR predictive modeling problem is formulated as ∗ ∗ and fqoe , which minimize their learning two functions, ftraf respective losses: ∗ ftraf = arg min Ltraf , ftraf

∗ fqoe = arg min Lqoe . fqoe

Fig. 1. System model of ResLearn-XR framework and XR Traffic-QoE platform.

IV. R ESIDUAL L EARNING F RAMEWORK As illustrated in Fig. 1, the ResLearn-XR platform interconnects XR head-mounted displays, wireless access networks, and cloud-rendering servers. Bidirectional IP traffic is passively captured per session as packet captures (pcap) traces, and user-reported session-level QoE labels (comfort, tolerable, discomfort, cybersickness) are associated with the corresponding traffic traces. On the traffic prediction path, frame count, average frame size, and inter-arrival interval are derived from packet-level observations, capturing delivery rate, buffering delay, and jitter, respectively, yielding the m = 3 targets for (1). On the QoE risk estimation path, the same

traces are processed by the DDA, which derives frame-timingaware descriptors from application-layer observables (packet length, direction, inter-arrival time, timestamp) aligned with the session’s configured refresh rate, enabling both tasks from a single shared traffic source. A. XR Traffic Prediction ResLearn-XR performs XR traffic prediction with a twostage model that forecasts frame-aligned traffic metrics derived from packet-level observations and relevant to the MTP latency constraint. These metrics aggregate packet measurements over temporal intervals synchronized with the XR

4

frame-generation period, so that each feature vector corresponds to one rendered-frame interval. The selected metrics— average frame size, frame count, and average frame interarrival interval—capture traffic volume, generation rate, and timing, respectively, which are the traffic dimensions most directly related to end-to-end latency. The segmented data preparation module in Fig. 1 uses the Frame Identification Algorithm (FIA) [26] to aggregate packetlevel network data over short, non-overlapping temporal windows: zk = FIA({xt | t ∈ Wk }) , (4) where Wk = [(k − 1)W, kW ) denotes the k-th window of duration W , and FIA(·) is the frame identification and aggregation operator. Applying FIA over an XR session yields Z = {zk ∈ R3 | k = 1, 2, . . . , K},

(5)

where each vector zk = [ sk , ck , ιk ] contains the average frame size sk , frame count ck , and average frame inter-arrival interval ιk for the corresponding window. Let ck denote the number of XR frames identified by FIA in window Wk . For each frame n ∈ {1, . . . , ck }, FIA estimates the frame size Sk,n from the aggregate payload size and the delivery timestamp τk,n from packet-level timestamps. The average frame size and average inter-arrival interval are then ( 1 Pck n=1 Sk,n , ck > 0, ck (6) sk = 0, ck = 0, ( 1 Pck
n=2 τk,n − τk,n−1 , ck > 1, ck −1 (7) ιk = 0, ck ≤ 1. The traffic prediction function in (1) is instantiated as a two-stage residual model, (traf)

ftraf = fθ

(traf)

+ gϕ

,

(8)

(traf)

where fθ denotes the base traffic prediction function and (traf) gϕ denotes the residual correction function. Given a lookback length S, the rolling input sequence is (S)

Zk

= [ zk−S+1 , . . . , zk ] ∈ RS×3 ,

(9)

where zk denotes the frame-aligned traffic vector for the k-th window, and the prediction target is the next-step traffic state yk+1 = zk+1 . The base predictor is   (base) (traf) (S) bk+1 = fθ y Zk , (10) (traf)

where fθ is parameterized by θ. ResLearn-XR uses a Transformer encoder as the default base model to capture longrange dependencies through self-attention, but the residual formulation is model-agnostic and can also be instantiated with LSTM, GRU, or Stacked LSTM encoders for controlled comparison. The base predictor is trained with the mean-squared error (MSE) objective   2 (base) (base) bk+1 Ltraf = E yk+1 − y , (11) 2

(base)

bk+1 is the base prediction, yk+1 is the FIA-derived where y ground-truth traffic vector, and the expectation is taken over the empirical training distribution. Although the MSE objective provides stable regression, it biases the base predictor toward conditional-mean behavior and attenuates short-term burst dynamics. Bursts are characterized by increased average frame size sk and frame count ck , often with reduced inter-arrival interval ιk ; underestimating these patterns weakens proactive congestion management and increases MTP-latency risk. As motivated in Appendix A, ResLearn-XR therefore introduces a value-space (traf) residual learner gϕ to model the remaining prediction error in the second stage of (1). With look-back length S, the first base prediction is generated at k = S: (base)

(traf)

bS+1 = fθ y

(S)

(ZS ).

After observing yS+1 , the first residual becomes available: (base)

bS+1 . eS+1 = yS+1 − y Consequently, residual refinement begins with the prediction of yS+2 , whereas yS+1 is predicted by the base model during residual warm-up. For any subsequent step, the base residual is defined after the corresponding ground-truth traffic state is observed: (base) bk+1 ∈ Rm , ek+1 = yk+1 − y (12) where m = 3 corresponds to the traffic vector [sk , ck , ιk ]. Thus, ek+1 is unavailable when predicting yk+1 . To preserve temporal causality, the residual learner uses the most recently observed residual, setting the residual look-back to Sr = 1; it receives ẽk as input and predicts the next correction b rk+1 . This one-step residual history avoids undefined early residual entries and maximizes the number of valid residual-training samples for short segmented traffic windows. To form non-negative residual inputs, we compute a global scalar bias from valid training residuals. Let Itr denote the training residual indices with j ≥ S + 1. The bias is   min ej,i ≥ 0. (13) b = min 0, j∈Itr , i∈{1,...,m}

The scalar b is computed once from the training residuals and remains fixed during validation, testing, and online inference. The bias-shifted residual is ẽj = ej + b1m ,

j = S + 1, . . . , K.

(14)

This shift preserves the temporal ordering of residuals and reduces mixed-sign cancellation when the residual learner models systematic underprediction during burst-dominated periods. Because the supervised target remains ek+1 , the predicted correction is expressed in the original residual space before being added to the base prediction. The residual learner estimates the next-step correction as (traf)

b rk+1 = gϕ

(ẽk ),

k ≥ S + 1.

(15)

5

At time k, ẽk is observed, whereas ek+1 becomes available Algorithm 1 ResLearn-XR: Two-Stage Traffic Prediction only after yk+1 is observed. With the base predictor fixed, the Require: Packet sequence X, window duration W , base lookback S, residual look-back Sr = 1 residual learner is trained by minimizing   bk+1 Ensure: Next-step traffic prediction y 2 (res) (traf) Ltraf = E ek+1 − gϕ (ẽk ) , k = S + 1, . . . , K − 1. 1: for k = 1, . . . , K do 2 2: zk ← FIA({xt }t∈Wk ) (16) 3: yk ← zk After residual warm-up, the final next-step traffic prediction 4: end for is (base) 5: Base-stage training: bk+1 = y bk+1 + b y rk+1 , k ≥ S + 1. (17) 6: for k = S, . . . , K − 1 do (S) For k = S, no previously observed residual exists, so the 7: Zk ← [zk−S+1 , . . . , zk ] (base) (traf) (S) prediction is base-only: bk+1 ← fθ 8: y (Zk ) (base) (base) bk+1 9: ek+1 ← yk+1 − y bS+1 = y bS+1 . y 10: end for Algorithm 1 summarizes the traffic prediction procedure, 11: Compute b once from valid training residuals: where W is the non-overlapping window duration, S is the   base-predictor look-back length, and Sr = 1 is the residual min ej,i . b = min 0, j∈Itr , i∈{1,...,m} look-back used by the value-space residual learner. The source 4 code of the XR traffic prediction module is public . 12: for j = S + 1, . . . , K do For a general residual look-back Sr > 1, residual-stage 13: ẽj ← ej + b1m training begins at k = S + Sr , because the earliest available 14: end for bias-shifted residual is ẽS+1 ; with Sr = 1, training begins at 15: Residual-stage training: k = S + 1. 16: for k = S + 1, . . . , K − 1 do (traf) 17: Train gϕ using ẽk as input and ek+1 as target 18: end for B. XR QoE Risk Estimation 19: Inference at time k: 1) DDA: The XR QoE risk estimation module estimates (S) 20: Zk ← [zk−S+1 , . . . , zk ] cybersickness risk from causal DDA descriptors. DDA first (base) (traf) (S) bk+1 ← fθ (Zk ) aggregates packet-level observables (packet size, direction, 21: y inter-arrival time, and timestamp) into frame-aligned statistics, 22: if k = S then (base) bk+1 ← y bk+1 y and then maps them to temporal descriptors. These descriptors 23: 24: else encode frame-rate deviation from the display target, timing and (traf) b rk+1 ← gϕ (ẽk ) throughput instability, and offered load as a congestion-delay 25: (base) bk+1 ← y bk+1 + b 26: y rk+1 proxy for supervised QoE risk estimation. DDA produces one descriptor per non-overlapping temporal 27: end if (base) bk+1 window Wk = [(k−1)W, kW ), k = 1, . . . , T . Each descriptor 28: After yk+1 is observed, compute ek+1 = yk+1 − y and update the residual state for the next prediction. is causal, using only packet observations with timestamps up bk+1 return y to kW ; rolling statistics use a fixed look-back of U descriptor indices, independent of W . k For window k, FIA identifies ck frames with sizes {Sk,i }ci=1 ck and inter-frame intervals {∆τk,i }i=2 , as in (6)–(7); sk and Inter-frame timing dispersion is q ιk follow (6) and (7). These FIA-derived quantities define
2 Pck  1 descriptors of frame delivery, timing instability, and load i=2 ∆τk,i − ιk , ck > 1, ck −1 jk = (20) conditions relevant to QoE risk. 0, ck ≤ 1. The effective frame delivery rate is Together, {fk , sk , ιk , rk , jk } causally characterize delivery ck , (18) rate, payload intensity, pacing regularity, offered load, and fk = W where deviations from the display target indicate under- timing variability, whose deviations increase MTP-latency and QoE-degradation risk. delivery and irregular frame pacing. Let R denote the candidate refresh-rate set; in XR sessions, The offered load is R = {60, 90, 120} Hz. Because the display refresh target r̂ is ck X 8 not directly observable from network traces, DDA estimates it rk = Sk,i , (19) W · 106 i=1 as r̂ = arg min r − median(f1 , . . . , fT ) . (21) r∈R where Sk,i is the size (bytes) of the i-th frame; rk (Mbps) proxies congestion pressure and queueing-delay variability. For non-standard or adaptive devices, the same estimator can use a device-supported or empirically inferred refresh-rate 4 https://github.com/yoga-suhas-km/ResLearn candidate set.

6

Given r̂, pacing descriptors are fk , r̂

TABLE II C AUSAL DDA DESCRIPTORS USED FOR Q O E RISK ESTIMATION .

r̂ − fk , r̂

dk = max{r̂ − fk , 0}, (22) where ρk measures pacing alignment, hk relative shortfall, and dk non-negative under-delivery. For each scalar sequence qk ∈ {rk , fk , ιk }, DDA computes causal dispersion and trend statistics over the preceding U windows. The normalized variability statistic is ρk =

hk =

CoVU (qk ) =

Std(qk−U +1:k ) , Mean(qk−U +1:k ) + ε

(23)

where ε > 0 ensures numerical stability, and CoVU (qk ) gives scale-normalized short-term volatility. Temporal drift over the same horizon is qk − qk−U βU (qk ) = . (24) U This statistic captures persistent descriptor trends, such as increasing offered load or degraded frame pacing. Table II summarizes the DDA descriptors. The composite instability indicator is
Jk = tanh λ1 CoVU (ιk ) + λ2 CoVU (fk ) + λ3 hk , (25) where λ1 , λ2 , and λ3 are non-negative weights selected via validation. The terms represent inter-frame timing volatility, frame-rate volatility, and display-target deviation; tanh(·) bounds the indicator while preserving sensitivity to instability. The DDA output for each session is H = {hk ∈ R12 | k = 1, . . . , T },

(26)

where each hk concatenates the twelve features in Table II:  hk = fk , sk , ιk , jk , rk , r̂, ρk ,  (27) hk , dk , CoVU (qk ), βU (qk ), Jk , where qk ∈ {rk , fk , ιk } denotes the scalar sequence used for the rolling statistics. For look-back S, the causal QoE input is (S)

Hk

= [ hk−S+1 , . . . , hk ] ∈ RS×12 .

(28)

Section VI-C evaluates descriptor-group contributions through an ablation of base per-window, pacing, short-term stability, composite-instability, and full DDA representations. 2) XR QoE Risk Estimation: Following (2), ResLearn-XR instantiates QoE risk estimation as a two-stage logit-space model,   (qoe) (qoe) fqoe = σ fθ + gϕ , (29) (qoe)

where fθ maps DDA-derived descriptors to a base logit, (qoe) gϕ adds a logit-space residual correction, and σ(·) returns the elevated-risk probability. The QoE branch follows the same temporal-causality princi(S) ple as the traffic-prediction branch: Hk = [hk−S+1 , . . . , hk ] contains only observations available up to window k, and the residual head uses only the corresponding latent state sk at inference. Future descriptors and ground-truth QoE labels are excluded.

Symbol Meaning Role in QoE risk estimation A. Base per-window descriptors fk Effective frame delivery Reduced rate increases MTP viorate lation risk sk Average payload size Larger frames amplify queueing per frame sensitivity ιk Mean inter-frame inter- Irregular spacing degrades motion val stability jk Inter-frame timing dis- Elevated values indicate pacing persion instability rk Window-level offered Higher load increases delay varitraffic load ability risk B. Pacing relative to display target r̂ Inferred display refresh Reference timing anchor for pactarget ing alignment ρk Frame pacing ratio Measures alignment with display (fk /r̂) timing hk Normalized pacing Quantifies relative under-delivery shortfall of frames dk Non-negative pacing Captures sustained under-pacing deficit events C. Short-term temporal stability (over U windows) CoVU (·) Coefficient of variation Captures short-term timing and rate volatility βU (·) Finite-difference trend Detects persistent temporal drift D. Composite instability indicator Jk Composite instability Aggregates correlated timing and index pacing cues

(S)

Given Hk ∈ RS×12 defined in (28), the base-stage estimator produces a scalar logit   (base) (qoe) (S) ℓk = fθ Hk ∈ R, (30) with base probability   (base) (base) , pbk = σ ℓk

1 . (31) 1 + e−u The base estimator is architecture-agnostic and can be instantiated with LSTM or GRU encoders. It is trained using classweighted binary cross-entropy with logits,    (base) L(base) = E − β yk log σ ℓk qoe     (32) (base) , − (1 − yk ) log 1 − σ ℓk σ(u) =

where yk ∈ {0, 1} is the binary QoE risk label (1: elevated risk, 0: non-elevated risk), and β = N0 /N1 compensates for class imbalance using the numbers of non-elevated-risk and elevated-risk training samples. To capture short-duration QoE degradations, let sk denote the final-step latent state produced by the base QoE backbone (S) from Hk . The residual head maps sk to a bounded corrective logit δk ∈ [−n, n]: (res)

ℓk

(base)

= ℓk + δk ,   (res) (res) pbk = σ ℓk .

(qoe)

δk = gϕ

(sk ),

(33) (34)

(qoe)

During residual-stage training, fθ is frozen and only (qoe) gϕ is optimized. The residual head is trained directly on the

7

corrected logit using the same class-weighted loss; no explicit residual target is required:    (base) L(res) + δk qoe = E − β yk log σ ℓk     (35) (base) − (1 − yk ) log 1 − σ ℓk + δk , where yk and β are defined as in (32). For binary QoE risk estimation, the predictor outputs     (res) fqoe = pbk , 1 − pbk , pbk = σ ℓk , (36) where pbk is the uncalibrated elevated-risk probability. Post-training calibration uses stratum-wise isotonic regression [27] on a held-out split disjoint from base and residual training. Strata depend only on DDA-derived covariates: the inferred refresh anchor r̂, offered load rk , pacing ratio ρk , and composite instability index Jk . Quantile discretizers Br , Bρ , and BJ partition rk , ρk , and Jk into Kr , Kρ , and KJ bins, inducing the calibration strata n o K = (r̂, br , bρ , bJ ) : r̂ ∈ R, br ∈ Br , . (37) bρ ∈ Bρ , bJ ∈ BJ Each window k is assigned by   κ(k) = Φ r̂, Br (rk ), Bρ (ρk ), BJ (Jk ) ∈ K,

Fig. 2. Flow of the XR QoE risk estimation module.

(38)

which groups samples with comparable load, pacing, and instability conditions. For stratum κ ∈ K, let Dκ collect calibration samples with κ(k) = κ. If non-empty, a stratum-specific isotonic map Cκ : [0, 1] → [0, 1] converts uncalibrated risk probabilities to calibrated probabilities while preserving monotonicity with empirical event frequency. If a fine-grained stratum is empty or unavailable, calibration backs off by progressively pooling conditioning variables:   C(r̂, br , bρ , bJ ) , D(r̂, br , bρ , bJ ) ̸= ∅,     C(r̂, br , bρ , ⋆) , D(r̂, br , bρ , ⋆) ̸= ∅,  (39) Cκ↑ = C(r̂, br , ⋆, ⋆) , D(r̂, br , ⋆, ⋆) ̸= ∅,    C(r̂, ⋆, ⋆, ⋆) , D(r̂, ⋆, ⋆, ⋆) ̸= ∅,    C , otherwise, global where “⋆” denotes pooling over that dimension; the most specific available calibrator is used, with Cglobal as the final fallback. (S) ↑ At inference, p̃k = Cκ(k) (b pk ) and fqoe (Hk ) = [p̃k , 1 − p̃k ]. The pipeline is shown in Fig. 2 and summarized in Algorithm 2; the implementation is public5 . Detailed implementation and reproducibility settings are given in Appendix B. V. E XPERIMENTATION S ETUP A. Experimental Platform and Dataset Collection The experimental platform consists of a Meta Oculus Quest 2 head-mounted display (HMD), an Xbox gaming console, a programmable traffic shaping module, and a remote XR content rendering system, as illustrated in Fig. 3. The Quest 2 5 https://github.com/yoga-suhas-km/XR QoE Prediction

Algorithm 2 ResLearn-XR: QoE Risk Estimation 1: Input: {hk }T k=1 , S, U = 16 2: Output: Calibrated risk vector [ p̃k , 1 − p̃k ] 3: for k = S, . . . , T do (S) 4: Hk ← [hk−S+1 , . . . , hk ] ▷ Causal look-back 5: end for 6: Base-stage training: 7: for k = S, . . . , T do (base) (qoe) (S) 8: ℓk ← fθ (Hk ) ▷ Base logit 9: end for (qoe) 10: Train fθ using class-weighted BCE-with-logits 11: Freeze θ 12: Residual-stage training: 13: for k = S, . . . , T do 14: Obtain final-step latent state sk from the frozen base QoE backbone ▷ Causal latent state (qoe) 15: δk ← gϕ (sk ) ▷ Bounded logit correction 16: end for (qoe) 17: Train gϕ using class-weighted BCE-with-logits on (base) ℓk + δk 18: Calibration: 19: for k = S, . . . , T do (base) 20: pbk ← σ(ℓk + δk ) ▷ Uncalibrated QoE risk probability
21: κ(k) ← Φ r̂, Br (rk ), Bρ (ρk ), BJ (Jk ) ▷ Calibration stratum 22: end for ↑ 23: Fit Cκ with hierarchical back-off to obtain Cκ 24: Inference at window k: (base) 25: pbk ← σ(ℓk + δk ) ↑ 26: p̃k ← Cκ(k) (b pk ) 27: return [ p̃k , 1 − p̃k ]

8

by assigning moderate/discomfort and severe/cybersickness sessions to the elevated-risk class (y = 1), and none/comfortable and slight/tolerable sessions to the nonelevated-risk class (y = 0). This protocol reduces participant burden and preserves immersion, but provides weak sessionlevel supervision rather than exact per-window discomfort onset, duration, or intensity. Accordingly, the QoE branch estimates whether traffic windows are associated with elevated session-level QoE risk, not directly measured instantaneous QoE.

Fig. 3. Experimental platform for creating XR Traffic-QoE dataset

HMD and Xbox console connect to a local router providing IP connectivity to the Internet, with network access mediated by a traffic shaping module that enforces controlled uplink and downlink rate constraints to emulate bandwidth-limited access network conditions relevant to XR QoE evaluation. The traffic shaping module is implemented on an Ubuntu 18.04 machine using the Wondershaper6 and is connected to the university backbone network, which offers an average unconstrained access bandwidth of approximately 120 Mbps. Remote XR rendering is performed on a cloud-hosted compute instance provided by Paperspace7 , configured with the Virtual Desktop Streamer (VDS) server application8 . The corresponding VDS client executes on the Quest 2 HMD and establishes a bidirectional streaming session with the remote rendering instance, such that XR application traffic traverses the Internet between the HMD and the cloud server. Data are captured from a diverse set of commercially available XR services spanning multiple modalities, including VR gaming (Hellblade: Senua’s Sacrifice VR Edition, DiRT Rally 2.0 VR), VR video streaming (Bigscreen VR), social VR with voice communication (VRChat), augmented reality (The Lab—Solar System), and mixed reality (Reality Mixer). Data collection is conducted with two participants to obtain temporally continuous XR traffic measurements and corresponding session-level QoE labels under controlled network conditions. For each participant, XR sessions are recorded across five bandwidth regimes (15, 30, 60, 120 Mbps, and adaptive), yielding repeated session measurements under identical application and network configurations. After each session, participants provide a four-level ordinal QoE rating aligned with ITU-T P.812/G.1035 and established cybersickness instruments [10], [11], [28]–[30]: none/comfortable, slight/tolerable, moderate/discomfort, and severe/cybersickness. These session-level ratings are binarized 6 https://github.com/magnific0/wondershaper 7 https://www.paperspace.com/ 8 https://www.vrdesktop.net/

B. Dataset and Experiments Table III summarizes the evaluation datasets. XR traffic prediction uses Dataset I [31] (in-house), Dataset II [14] (SteamVR traces: SteamVR Home and Beat Saber), and Dataset III [17] (two subsets: diverse VR applications, and a replay under varied network conditions). Dataset I also supports QoE risk estimation, as no public dataset provides paired continuous traffic traces and session-level QoE labels; its details are in Section V-A. Traffic prediction follows the effective chronological 40%/10%/50% train/validation/test split within each continuous XR traffic session. Input–target sequences are generated separately within each segment, without shuffling or splitboundary crossing, to prevent look-ahead leakage and evaluate next-step prediction on later unseen traffic. For QoE risk estimation, splitting is performed at the session level because all windows from a session share a weak QoE label. Consistent with Tables XII–XIV, we report participant-independent evaluation, which trains and tests on complete sessions from different participants, and mixeduser evaluation, which assigns complete sessions from both participants to disjoint training, validation, and testing sets. This design prevents leakage of session-specific patterns and repeated labels; no separate participant-dependent QoE setting is reported. For Dataset I, all five bandwidth regimes (15, 30, 60, 120 Mbps, and adaptive) are represented in the evaluated splits. Traffic prediction and mixed-user QoE splits are bandwidth-stratified where possible, whereas participantindependent QoE evaluation separates users while retaining matched bandwidth regimes for both participants. This supports evaluation under heterogeneous access-rate conditions and limits dominance by any single bandwidth setting. Informer, FEDformer, and Temporal Fusion Transformer are included as strong traffic-forecasting baselines, while logistic regression, random forest, and gradient-boosted trees serve as non-deep-learning QoE baselines. All models use the corresponding task-specific inputs, splits, and metrics. The framework is implemented in Python (NumPy [32], Pandas [33], Scikit-learn [34], TensorFlow [35], PyTorch [36]) and trained on an NVIDIA RTX 2080 Super GPU. C. Performance Evaluation Metrics 1) Traffic Prediction: XR traffic prediction performance is evaluated using Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Symmetric Mean Absolute Percentage Error (SMAPE) [6]. These metrics measure absolute and relative deviations between predicted and observed traffic values.

9

TABLE III DATASETS AND E VALUATION S ETTINGS FOR XR T RAFFIC AND Q O E R ISK P REDICTION

D-II

Exp 1 Exp 1

III

D-I

Exp.

Exp 1

D-I

Exp 2

Exp 1

Services Traffic Prediction VR Game, VR Video, VR Chat/VoIP, AR, MR

Applications DiRT Rally 2.0, Bigscreen VR, VR Chat, Solar System, Reality Mixer

Slow VR Traffic, Fast VR Steam VR Home, Beat Saber Traffic Fast VR Game 1, Fast VR Beat Saber, Medal of Honor, Game 2, Slow VR Game 1, Forklift Sim, Cooking Sim Slow VR Game 2 Slow VR Traffic, Fast VR Forklift Sim, Cooking Sim, Traffic Beat Saber, Medal of Honor QoE Risk Prediction DiRT Rally 2.0, Bigscreen VR, VR Game, VR Video, VR VR Chat, Solar System, Chat/VoIP, AR, MR Reality Mixer

TABLE IV F RAME - COUNT PREDICTION PERFORMANCE ON DATASET II, COMPARING BASELINE TEMPORAL MODELS AND THEIR R ES L EARN -XR VARIANTS .

TABLE V F RAME - SIZE PREDICTION PERFORMANCE ON DATASET I, COMPARING BASELINE TEMPORAL MODELS AND THEIR RESLEARN - XR VARIANTS . Model Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Model

Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Causal SMAPE Gain

– – – – 8.48% 6.49% 3.37% 7.69%

2) QoE Risk Estimation: QoE risk estimation performance is evaluated using SMAPE, Expected Calibration Error (ECE) [37], and Area Under the Receiver Operating Characteristic Curve (AUC) [38]. For the DDA descriptor ablation, we additionally report Quadratic Weighted Kappa (QWK) and Macro-F1. QWK measures agreement while accounting for the ordinal structure of QoE labels, Macro-F1 evaluates class-balanced prediction performance, and ECE evaluates probability calibration. VI. R ESULTS AND D ISCUSSIONS A. Traffic Prediction Tables IV–VII report the corrected strictly causal trafficprediction results. Across 16 paired backbone–task comparisons, every ResLearn-XR variant reduces SMAPE relative to its base temporal model, with gains from 0.35% to 17.84% and an average gain of approximately 6.7%. These results support residual correction as a consistent but task-dependent causal refinement of the base forecasters. The gains vary by traffic descriptor and backbone. Table IV shows frame-count SMAPE reductions of 3.37%–8.48% on Dataset II. For frame-size prediction, Tables V and VI show maximum gains of 10.65% and 14.21%, respectively; in Dataset III Experiment 2, the LSTM residual variant gives the lowest absolute error, whereas the Transformer yields the largest relative gain. Table VII reports the largest overall improvement, 17.84%, for Transformer-based inter-arrival-time

Causal SMAPE Gain – – – – 10.65% 6.49% 7.54% 2.27%

TABLE VI F RAME - SIZE PREDICTION PERFORMANCE ON DATASET III (E XPERIMENT 2), COMPARING BASELINE TEMPORAL MODELS AND THEIR RESLEARN - XR VARIANTS . Model

Metrics RMSE MAPE SMAPE Base temporal models 0.0427 0.0074 0.7453 0.0420 0.0074 0.7412 0.0420 0.0074 0.7412 0.0425 0.0075 0.7516 ResLearn-XR 0.0385 0.0068 0.6821 0.0389 0.0069 0.6931 0.0417 0.0071 0.7162 0.0390 0.0069 0.6938

Metrics RMSE MAPE SMAPE Base temporal models 5294.76 0.0085 0.8506 5103.71 0.0082 0.8195 5318.15 0.0086 0.8619 5104.09 0.0082 0.8222 ResLearn-XR 4760.99 0.0076 0.7600 4788.68 0.0076 0.7663 4953.34 0.0079 0.7968 4987.10 0.0080 0.8035

Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Metrics MAPE SMAPE Base temporal models 53573.43 0.0051 0.5094 45694.86 0.0045 0.4445 50025.70 0.0046 0.4608 46472.13 0.0043 0.4258 ResLearn-XR 47578.00 0.0044 0.4370 43514.14 0.0040 0.4036 48598.58 0.0045 0.4498 44303.85 0.0042 0.4232 RMSE

Causal SMAPE Gain

– – – – 14.21% 9.20% 2.39% 0.61%

prediction, while the GRU gain is marginal at 0.35%. Thus, residual learning is best interpreted as a lightweight, taskdependent refinement mechanism. Table VIII compares the Transformer-based ResLearn-XR model with stronger forecasting baselines under the same chronological split. ResLearn-XR attains the lowest RMSE, MAPE, and SMAPE, improving the best competing values from 4818.62 to 4760.99, 0.0085 to 0.0076, and 0.85 to 0.76, respectively. The controlled residual-learning claim remains based on the paired comparisons in Tables IV–VII. Using the queueing proxy in Appendix A, the predicted bk = b offered load is estimated as L ck sbk /(b ιk + ϵ) and marked as MTP-risk positive when it exceeds a predefined threshold.

TABLE VII F RAME INTER - ARRIVAL TIME PREDICTION PERFORMANCE ON DATASET III (E XPERIMENT 1), COMPARING BASELINE TEMPORAL MODELS AND THEIR RESLEARN - XR VARIANTS . Model

Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Metrics MAPE SMAPE Base temporal models 0.1371 0.0118 1.1752 0.1425 0.0121 1.2032 0.1309 0.0114 1.1351 0.1393 0.0118 1.1735 ResLearn-XR 0.1248 0.0106 0.9655 0.1303 0.0113 1.1264 0.1307 0.0113 1.1311 0.1322 0.0113 1.1312

RMSE

Causal SMAPE Gain

– – – – 17.84% 6.38% 0.35% 3.60%

10

TABLE VIII C OMPARISON WITH T RANSFORMER - BASED FORECASTING BASELINES . Model Informer FEDformer Temporal Fusion Transformer ResLearn-XR, Transformer backbone

RMSE 4818.62 5676.41 4958.27 4760.99

MAPE 0.0093 0.0092 0.0085 0.0076

SMAPE 0.92 0.93 0.85 0.76

TABLE IX A PPROXIMATE MTP- RISK DETECTION ANALYSIS USING A QUEUEING - DELAY PROXY DERIVED FROM PREDICTED BURST TRAFFIC . Model Base-only predictor Base + residual predictor

Precision 0.78 0.84

Recall 0.63 0.81

F1 0.70 0.82

Missed-Risk Rate 0.37 0.19

Table IX shows that residual correction increases proxy-risk recall from 0.63 to 0.81 and reduces the missed-risk rate from 0.37 to 0.19. This indicates improved detection of burstinduced risk windows, while direct closed-loop MTP-violation measurement remains future work. Although this study focuses on predictive modeling rather than closed-loop control, ResLearn-XR outputs can support XR-aware management. Traffic predictions provide offeredload and burst-risk indicators for proactive bandwidth, queue, or rate-adaptation decisions, while calibrated QoE-risk probabilities can inform admission control, edge scaling, rendering adaptation, or migration. The bandwidth settings used here define controlled evaluation conditions and are not dynamically adjusted; closed-loop resource orchestration remains future work. Table X reports approximate trainable parameters, FP32 weight memory, and per-window inference latency. The residual head adds only about 4K parameters and 0.02 MB to each backbone; for the Transformer, latency increases from 0.42 ms to 0.47 ms, and all recurrent variants remain below 1 ms. These values support low-latency model-only inference, but they do not include packet capture, descriptor extraction, controller communication, scheduling, or actuation delay; therefore, full deployment still requires system-level profiling and closedloop validation. B. QoE Risk Estimation QoE risk estimation evaluates probabilistic prediction of the elevated-risk event QoE ≤ 2 under user and network variability. Tables XII–XIV and Fig. 4 summarize prediction error, calibration, discrimination, and temporal risk behavior. Under participant-independent evaluation, ResLearn-XR reduces SMAPE for all reported backbones. The Transformer residual model decreases SMAPE from 1.250 to 0.259 for U2→U1 and from 1.245 to 0.152 for U1→U2, while increasing AUC to 0.874 and 0.930, respectively. In the mixeduser setting, residual models also reduce SMAPE; the LSTM variant gives the lowest error (0.429), while the Transformer gives the highest AUC (0.909). Calibration is split- and backbone-dependent. For example, in Table XIII, the non-residual Transformer and Stacked LSTM have lower ECE (0.052) than the residual Transformer

TABLE X A PPROXIMATE MODEL COMPLEXITY, FP32 TRAINABLE - WEIGHT MEMORY, AND MODEL - ONLY FEASIBILITY COMPARISON . Model

Trainable Parameters

FP32 Weight Inference Time / Memory Window

Modelonly Feasible Yes Yes

Transformer ∼265K ∼1.06 MB ∼0.42 ms Transformer + ∼269K ∼1.08 MB ∼0.47 ms Residual LSTM ∼68K ∼0.27 MB ∼0.31 ms Yes LSTM + Residual ∼72K ∼0.29 MB ∼0.36 ms Yes GRU ∼51K ∼0.20 MB ∼0.28 ms Yes GRU + Residual ∼55K ∼0.22 MB ∼0.33 ms Yes Stacked LSTM ∼200K ∼0.80 MB ∼0.55 ms Yes Stacked LSTM + ∼204K ∼0.82 MB ∼0.61 ms Yes Residual Note: Feasibility refers to model-only inference and excludes packet capture, descriptor extraction, controller communication, scheduling, and actuation delay.

TABLE XI C OMPARISON WITH NON - DEEP - LEARNING Q O E RISK BASELINES UNDER PARTICIPANT- INDEPENDENT EVALUATION . Model Logistic regression Random forest Gradient-boosted trees ResLearn-XR

AUC 0.681 0.754 0.812 0.930

ECE 0.146 0.113 0.089 0.057

Macro-F1 0.382 0.421 0.458 0.497

(0.076), whereas the residual Transformer improves ECE in Table XII. Therefore, SMAPE, AUC, and ECE are interpreted as complementary rather than uniformly aligned model properties. Figure 4 qualitatively supports these trends. ResLearn-XR produces sustained high-risk trajectories for sessions associated with cybersickness and low-variance trajectories for comfortable sessions, whereas the displayed non-residual baselines are noisier and less separable. Because the QoE dataset contains two participants and session-level weak labels, the participant-independent results should be interpreted as preliminary cross-subject evidence rather than broad user-level generalization. The results support the feasibility of traffic-based QoE risk estimation under weak supervision, while larger cohorts, temporally localized feedback, and broader XR scenarios are required for robust user-independent validation. Table XI shows that ResLearn-XR attains the highest AUC and Macro-F1 and the lowest ECE among the non-deeplearning baselines, indicating that temporal residual modeling adds value beyond the DDA-derived descriptors alone. C. Effect of the DDA Descriptors To evaluate the contribution of the proposed DDA, we compare QoE risk estimation performance with base perwindow features alone and with the full DDA descriptor set. Results are reported in Table XV for the participantindependent configuration trained on User 1 and evaluated on User 2. Incorporating DDA descriptors yields consistent improvements across all evaluation metrics. Relative to the base perwindow features, SMAPE decreases from 0.2831 to 0.1524,

11

TABLE XII Q O E RISK ESTIMATION PERFORMANCE UNDER PARTICIPANT- INDEPENDENT EVALUATION (U2 TRAIN / U1 TEST ).

Model Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Metrics SMAPE ECE AUC % SMAPE Improvement Non-ResLearn-XR Algorithm 1.250 0.283 0.764 1.259 0.170 0.679 1.114 0.462 0.571 1.167 0.269 0.752 ResLearn-XR Solution 0.259 0.119 0.874 79.28% 0.662 0.329 0.679 47.42% 0.923 0.267 0.571 17.15% 0.528 0.247 0.752 54.76%

(a) DiRTRally 2.0 (15 Mbps, 60 Hz) (b) RealityMixer (30 Mbps, 120 Hz)

(c) The Lab (AB cloud, 60 Hz)

(d) Bigscreen (120 Mbps, 90 Hz)

(e) Hellblade (30 Mbps, 120 Hz)

(f) The Lab (60 Mbps, 90 Hz)

TABLE XIII Q O E RISK ESTIMATION PERFORMANCE UNDER PARTICIPANT- INDEPENDENT EVALUATION (U1 TRAIN / U2 TEST ).

Model Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Metrics SMAPE ECE AUC % SMAPE Improvement Non-ResLearn-XR Algorithm 1.245 0.052 0.500 1.068 0.084 0.500 1.347 0.496 0.390 1.245 0.052 0.500 ResLearn-XR Solution 0.152 0.076 0.930 87.79% 0.507 0.253 0.706 52.53% 0.369 0.154 0.821 72.61% 0.709 0.354 0.539 43.05%

ECE decreases from 0.1399 to 0.0760, and AUC increases from 0.8607 to 0.9302. These gains indicate that the DDA descriptors improve not only prediction accuracy but also probability calibration and discriminative capability under cross-user generalization. The observed improvements can be attributed to the structured, causally motivated design of the DDA descriptors, which encode timing alignment, headroom, and short-term traffic variability information derived from application-layer observables. By augmenting base-per-window statistics with descriptors explicitly aligned with XR system timing constraints, the model receives richer contextual information for QoE risk estimation. While Table XV does not isolate the conTABLE XIV Q O E RISK ESTIMATION PERFORMANCE UNDER MIXED - USER EVALUATION .

Model Transformer LSTM GRU Stacked LSTM Transformer LSTM GRU Stacked LSTM

Metrics SMAPE ECE AUC % SMAPE Improvement Non-ResLearn-XR Algorithm 1.279 0.296 0.563 1.211 0.365 0.656 0.603 0.258 0.760 1.216 0.241 0.612 ResLearn-XR Solution 0.817 0.179 0.909 36.12% 0.429 0.214 0.802 64.57% 0.500 0.168 0.875 17.08% 1.167 0.357 0.802 4.03%

Fig. 4. Predicted QoE risk trajectories p(QoE ≤ 2) for representative XR sessions under varying applications and network conditions. (a), (b), and (e) correspond to sessions associated with cybersickness, while (c), (d), and (f) correspond to low-risk sessions. Trajectories in (a)–(d) are produced by ResLearn-XR, whereas (e) and (f) are produced by non-ResLearn-XR baseline models.

tribution of individual descriptors, the aggregate performance gains demonstrate the effectiveness of the DDA formulation as a whole in improving robustness and calibration across users. TABLE XV E FFECT OF DDA DESCRIPTORS ON Q O E RISK ESTIMATION PERFORMANCE UNDER PARTICIPANT- INDEPENDENT EVALUATION (U1 TRAIN / U2 TEST ).

Feature Set Base per-window features DDA features

SMAPE ↓ 0.2831 0.1524

ECE ↓ 0.1399 0.0760

AUC ↑ 0.8607 0.9302

Table XVI reports an ablation study of the DDA descriptor groups. The base per-window descriptors provide the fundamental traffic state, while pacing-related descriptors characterize alignment with the inferred display refresh target. The short-term stability descriptors capture temporal variability and drift, and the composite instability index aggregates timing variability and pacing shortfall into a bounded QoE-sensitive descriptor. Among the evaluated feature groups, adding the composite instability index provides the largest gain, achieving the highest QWK and Macro-F1 and the lowest ECE. The full DDA representation also improves calibration compared with the base descriptor set, reducing ECE from 0.1413 to 0.1196 while maintaining strong agreement. These results indicate that the DDA performance is mainly driven by instabilityaware temporal descriptors, which capture QoE-relevant traffic irregularities beyond base per-window traffic statistics.

12

TABLE XVI A BLATION STUDY OF DDA DESCRIPTOR GROUPS FOR Q O E ESTIMATION . Feature Set Base per-window descriptors Base + pacing-related descriptors Base + short-term stability descriptors Base + composite instability index Full DDA

QWK 0.7999 0.7676 0.7938 0.8219 0.8064

Macro-F1 0.5228 0.4940 0.4993 0.5242 0.4972

ECE 0.1413 0.1382 0.1300 0.1196 0.1196

TABLE XVII A BLATION OF THE RESIDUAL BIAS TERM FOR FRAME - SIZE PREDICTION ON DATASET I. R ESULTS ARE REPORTED AS MEAN ± STANDARD DEVIATION OVER FIVE INDEPENDENT RUNS . Residual configuration Without bias (b = 0) With training-set bias

RMSE 4908.67 ± 42.31 4764.95 ± 31.84

MAPE 0.0076 ± 0.00008 0.0054 ± 0.00006

SMAPE 0.97 ± 0.009 0.76 ± 0.007

D. Effect of Residual Learning For XR traffic prediction, residual learning improves shortterm deviations superimposed on longer temporal trends. Tables IV–VII show positive SMAPE gains for all backbone– task pairs, with the largest gain of 17.84% for Transformerbased inter-arrival-time prediction and smaller recurrent-model gains in some settings. The residual pathway is therefore best interpreted as a lightweight causal refinement layer for structured prediction errors. Table XVII shows that incorporating the training-set-derived residual bias improves frame-size prediction across all reported metrics. Relative to the zero-bias residual configuration, the training-set bias reduces RMSE from 4908.67 to 4764.95, corresponding to an approximate 2.9% reduction. MAPE decreases from 0.0076 to 0.0054, and SMAPE decreases from 0.97 to 0.76, corresponding to approximate reductions of 28.9% and 21.6%, respectively. The experiments are repeated over five independent runs, and the smaller standard deviations for the biased configuration indicate more stable performance across random initializations. Because the bias is computed only once from the training residuals and then fixed during validation, testing, and inference, these gains reflect correction of a systematic residual offset rather than leakage from future test samples. For QoE risk estimation, residual learning primarily improves predictive accuracy and discrimination, while calibration remains backbone- and split-dependent. The reliability diagrams in Fig. 5 provide an illustrative example in which the displayed ResLearn-XR model is better calibrated than the displayed non-residual baseline; ECE trends across all models are reported in Tables XII–XIV. E. Deployment Feasibility ResLearn-XR is intended as a lightweight online predictive component deployed at an edge server, access-network controller, or monitoring gateway. Its inputs are causally derived from packet timing, packet size, and packet direction, thus requiring no payload inspection or application-layer decryption. Table X shows that the Transformer-based ResLearnXR configuration adds only approximately 4K parameters,

(a) ResLearn-XR

(b) Non-ResLearn-XR

Fig. 5. Reliability comparison of QoE risk calibration on U2 test data. For the displayed model pair, the ResLearn-XR model in (a) aligns more closely with the diagonal than the non-ResLearn-XR baseline in (b).

0.02 MB of FP32 trainable-weight memory, and 0.05 ms per prediction window over the base Transformer (269K parameters, 1.08 MB FP32 trainable-weight memory, and 0.47 ms total), while the recurrent variants remain below 1 ms. These results support low-latency model-only inference, but they do not include packet capture, descriptor extraction, controller communication, scheduling, or actuation delay; therefore, full deployment still requires system-level profiling and closedloop validation. F. Limitations The main limitations are the small two-participant QoE dataset, session-level weak labels assigned to all temporal windows, and the binary reduction of four ordinal subjective ratings. As a result, the participant-independent results should be interpreted as preliminary cross-user evidence rather than broad user-level generalization, and the QoE model may capture participant-, application-, bandwidth-, or session-specific correlates. Broader validation with larger cohorts, temporally localized QoE annotations, additional devices and network conditions, and controller-in-the-loop experiments is required before making broad deployment claims. VII. C ONCLUSION AND F UTURE W ORK ResLearn-XR is a two-stage residual learning framework for cognitive XR network management, comprising value-space and logit-space residual heads for traffic prediction and QoE risk estimation, respectively. Grounded in ITU-T QoE causality via frame-timing-aware DDA descriptors, the corrected strictly causal traffic-prediction results show positive SMAPE gains up to 17.84% and an average gain of approximately 6.7% across the updated traffic tables, while QoE risk estimation achieves up to 87.8% SMAPE improvement over single-stage baselines. The current QoE evaluation is limited by the two-participant dataset and session-level weak labels, so the participantindependent results should be interpreted as preliminary crossuser evidence rather than broad user-level generalization. ResLearn-XR should be viewed as a predictive modeling component for supporting proactive bandwidth allocation,

13

scheduling, rate adaptation, admission control, and edgeresource orchestration, rather than as an end-to-end closedloop controller. Future work will expand the XR Traffic-QoE dataset and evaluate heterogeneous HMDs and 5G/6G access conditions, while controller-in-the-loop deployment is left for future work. A PPENDIX A XR T RAFFIC S PIKES AND MTP V IOLATION The MTP latency decomposes as LMTP (t) = Lsense +Lrender +Lenc +Lnet (t)+Ldec +Ldisplay , where sensing, rendering, encoding, decoding, and display delays are near-constant relative to network dynamics. Grouping them as Lproc , the ITU-T 20 ms constraint imposes a network latency budget M = 20 ms−Lproc , and MTP stability requires Lnet (t) ≤ M . Over a window of duration ∆W ≈ W ῑt , the total offered traffic is A(t) = ct st (bits), giving arrival rate st ct st ≈ (bits/s), λ(t) = ∆W ιt since ct ≈ ∆W /ιt , and utilization ρ(t) = λ(t)/C < 1. This shows that all three traffic metrics (st , ct , ιt ) jointly determine network load. Modeling the bottleneck as an M/M/1 queue [39] with service rate C, the expected queueing delay per bit is (bit) E[Wq (t)] = ρ(t)/[C(1 − ρ(t))], so the frame-level network delay is st Lnet (t) ≈ L0 + , C − λ(t) where L0 accounts for fixed propagation delays. As ρ(t) → 1− , differentiation yields (bit)

∂ E[Wq ] 1 = , ∂ρ(t) C(1 − ρ(t))2 showing that even small burst-driven increases in st or ct , or reductions in ιt , can exhaust M and violate the 20 ms constraint. For general traffic, Kingman’s approximation for a G/G/1 queue gives E[Wq(bit) (t)] ≈

ρ(t) c2a + c2s , 1 − ρ(t) 2C

where c2a and c2s are the squared coefficients of variation of inter-arrival and service times. The same (1 − ρ)−1 divergence persists, confirming that bursty XR traffic produces disproportionate latency increases and motivating explicit residual correction of short-horizon predictions of (st , ct , ιt ). This relationship also motivates the proxy-based MTP-risk analysis in Section VI-A, where each prediction window uses ck sbk bk = b , (40) L b ιk + ϵ with ϵ preventing division by zero. A burst-risk indicator is then b k > Ceff ), vbk = I(L (41) where Ceff is an effective service-capacity threshold. Residual correction reduces burst underestimation and improves proxyrisk detection, while direct closed-loop MTP-violation measurement remains future work.

A PPENDIX B I MPLEMENTATION AND R EPRODUCIBILITY C ONFIGURATION TABLE XVIII I MPLEMENTATION AND REPRODUCIBILITY CONFIGURATION FOR R ES L EARN -XR. Item

Configuration Traffic prediction branch Targets FIA-derived frame count, average frame size, and average frame inter-arrival time. Window duration W Dataset I: 1 s; Datasets II–III: 0.5 s. Rolling aggregation winFrame size/count/IAT windows: Dataset I dow 20/22/5; Dataset II Exp. 1 20/22/10; Dataset II Exp. 2–6 20/22/5; Dataset III Exp. 1–2 25/20/5. Base look-back S Dataset I: 17; Dataset II Exp. 1: 20; Dataset II Exp. 2–6: 5; Dataset III Exp. 1–2: 8 windows. Residual look-back Sr Sr = 1; only the latest observed bias-shifted residual ẽk is used to predict r̂k+1 . Training split Chronological within-session 40/10/50% train/validation/test split; sequences are constructed within each segment to avoid split-boundary leakage. (base) Base traffic loss MSE between ŷk+1 and yk+1 . (traf)

MSE between gϕ (ẽk ) and ek+1 . Adam in TensorFlow/Keras; optimizer states are excluded from Table X FP32 memory. Epochs and batch size 20 epochs; batch size 1. Weight decay None in the released traffic-prediction configuration. Random seed Fixed where specified; multi-run tables report mean ± standard deviation. Traffic prediction architectures Backbones Transformer, LSTM, GRU, and Stacked LSTM under the two-stage residual framework in Table X. Transformer Encoder with multi-head self-attention, feedforward projection, dropout, and dense output head. LSTM / GRU Single recurrent backbone and dense output head. Stacked LSTM Two-layer recurrent backbone with inter-layer dropout and dense output head. Residual traffic head Lightweight dense value-space correction head trained on the corrected causal residual target. Complexity accounting Table X counts trainable parameters for the active encoder, prediction head, and residual head; FP32 memory uses four bytes/weight and excludes activations, buffers, and optimizer states. Inference timing Single-window, batch-size-one latency after descriptor construction, compared with the 20 ms MTP budget; packet capture, control, scheduling, and actuation delays require separate profiling. QoE risk-estimation branch (S) Input descriptors Causal DDA history Hk = [hk−S+1 , . . . , hk ]. DDA look-back U Fixed causal horizon for CoVU (·) and βU (·): U = 16. QoE label Binary weak-supervision label from sessionlevel QoE annotation. Class weight β = N0 /N1 , using non-elevated-risk (N0 ) and elevated-risk (N1 ) training samples. QoE base loss Class-weighted binary cross-entropy with logits. QoE residual head Lightweight dense logit-space correction head applied before probability calibration. QoE residual loss Class-weighted binary cross-entropy on cor(base) rected logit ℓk + δk . Calibration split Dedicated split disjoint from base and residual training data. Calibration method Stratum-wise isotonic regression using DDAderived covariates only. Calibration strata Inferred refresh anchor r̂, offered-load bin Br (rk ), pacing-ratio bin Bρ (ρk ), and instability bin BJ (Jk ). Calibration back-off Use the most specific non-empty stratum calibrator; otherwise use the global isotonic calibrator. Residual traffic loss Optimizer

14

R EFERENCES [1] L. Zhang, X. Wu, F. Wang, A. Sun, L. Cui, and J. Liu, “Edge-Based Video Stream Generation for Multi-Party Mobile Augmented Reality,” IEEE Transactions on Mobile Computing, vol. 23, no. 1, pp. 409–422, 2024. [2] M. Chen, W. Saad, and C. Yin, “Virtual Reality Over Wireless Networks: Quality-of-Service Model and Learning-Based Resource Management,” IEEE Transactions on Communications, vol. 66, no. 11, pp. 5621–5635, 2018. [3] M. S. Elbamby, C. Perfecto, M. Bennis, and K. Doppler, “Toward LowLatency and Ultra-Reliable Virtual Reality,” IEEE Network, vol. 32, no. 2, pp. 78–84, 2018. [4] Y. Pei, M. Li, X. Huang, and X. Shen, “QoE-Aware Volumetric Video Caching and Rendering for Mobile Extended Reality Services,” IEEE Internet of Things Journal, vol. 12, no. 12, pp. 21 852–21 865, 2025. [5] B. Lim, S. Ö. Arık, N. Loeff, and T. Pfister, “Temporal Fusion Transformers for Interpretable Multi-Horizon Time Series Forecasting,” International Journal of Forecasting, vol. 37, no. 4, pp. 1748–1764, 2021. [6] S. Vaidya, H. Abou-Zeid, and D. Krishnamurthy, “Transfer Learning for Online Prediction of Virtual Reality Cloud Gaming Traffic,” in 2023 IEEE Global Communications Conference (GLOBECOM), 2023, pp. 4668–4673. [7] J. Woo, S. Hong, D. Kang, and D. An, “Improving the Quality of Experience of Video Streaming Through a Buffer-Based Adaptive Bitrate Algorithm and Gated Recurrent Unit-Based Network Bandwidth Prediction,” Applied Sciences, vol. 14, no. 22, p. 10490, 2024. [8] H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond Efficient Transformer for Long Sequence TimeSeries Forecasting,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115. [9] T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency Enhanced Decomposed Transformer for Long-Term Series Forecasting,” in International Conference on Machine Learning. PMLR, 2022, pp. 27 268–27 286. [10] International Telecommunication Union, “Framework of Quality of Experience for Virtual Reality Services,” ITU-T, Geneva, Switzerland, Tech. Rep. G.1035, 2020. [11] International Telecommunication Union (ITU), “Principles of Subjective Test Methods for Interactive Virtual Reality (VR) Applications,” ITU-T, Geneva, Switzerland, Recommendation, May 2024. [12] G. Kougioumtzidis, A. Vlahov, V. K. Poulkov, P. I. Lazaridis, and Z. D. Zaharis, “Deep Learning-Aided QoE Prediction for Virtual Reality Applications Over Open Radio Access Networks,” IEEE Access, vol. 11, pp. 143 514–143 529, 2023. [13] J. Yu, A. Y. Alhilal, T. Zhou, P. Hui, and D. H. K. Tsang, “AttentionBased QoE-Aware Digital Twin Empowered Edge Computing for Immersive Virtual Reality,” IEEE Transactions on Wireless Communications, vol. 23, no. 9, pp. 11 276–11 290, 2024. [14] S. Zhao, H. Abou-zeid, R. Atawia, Y. S. K. Manjunath, A. B. Sediq, and X.-P. Zhang, “Virtual Reality Gaming on the Cloud: A Reality Check,” in 2021 IEEE Global Communications Conference (GLOBECOM), 2021, pp. 1–6. [15] J. Wang, H.-N. Liang, D. Monteiro, W. Xu, and J. Xiao, “Real-Time Prediction of Simulator Sickness in Virtual Reality Games,” IEEE Transactions on Games, vol. 15, no. 2, pp. 252–261, 2023. [16] M. Polupanova, “VR Traffic Dataset on Broad Range of End-User Activities,” Data, vol. 8, no. 8, p. 132, 2023. [Online]. Available: https://www.mdpi.com/2306-5729/8/8/132 [17] S. Baldoni, F. Battisti, F. Chiariotti, F. Mistrorigo, A. B. Shofi, P. Testolina, A. Traspadini, A. Zanella, and M. Zorzi, “Questset: A VR Dataset for Network and Quality of Experience Studies,” in Proceedings of the 15th ACM Multimedia Systems Conference, 2024, pp. 408–414. [18] Y. S. K. Manjunath, M. Szymanowski, A. Wissborn, M. Li, L. Zhao, and X.-P. Zhang, “ResLearn: Transformer-based Residual Learning for Metaverse Network Traffic Prediction,” in 2025 IEEE International Conference on Communications (ICC). IEEE, 2025, pp. 5053–5058. [19] M. Lecci, M. Drago, A. Zanella, and M. Zorzi, “An Open Framework for Analyzing and Modeling XR Network Traffic,” IEEE Access, vol. 9, pp. 129 782–129 795, 2021. [20] D. G. Morı́n, D. Medda, A. Iossifides, P. Chatzimisios, A. G. Armada, A. Villegas, and P. Peréz, “An eXtended Reality Offloading IP Traffic Dataset and Models,” IEEE Transactions on Mobile Computing, vol. 23, no. 6, pp. 6820–6834, 2023.

[21] F. Chiariotti, M. Drago, P. Testolina, M. Lecci, A. Zanella, and M. Zorzi, “Temporal Characterization and Prediction of VR Traffic: A Network Slicing Use Case,” IEEE Transactions on Mobile Computing, vol. 23, no. 5, pp. 3890–3908, 2024. [22] J. Ruan and D. Xie, “A Survey on QoE-Oriented VR Video Streaming: Some Research Issues and Challenges,” Electronics, vol. 10, no. 17, p. 2155, 2021. [23] N. Liu, T. H. Luan, Y. Wang, Y. Liu, and Z. Su, “QoE-Oriented Cooperative VR Rendering and Dynamic Resource Leasing in Metaverse,” IEEE Transactions on Mobile Computing, 2025. [24] L. Huang, Y. Li, H. Liang, K. Chi, and Y. Wu, “Enhanced VR Experience with Edge Computing: The Impact of Decoding Latency,” IEEE Transactions on Mobile Computing, 2025. [25] R. D. Tripathi, M. Lyu, and V. Sivaraman, “Assessing the Impact of Network Quality-of-Service on Metaverse Virtual Reality User Experience,” in 2024 IEEE International Conference on Metaverse Computing, Networking, and Applications (MetaCom), 2024, pp. 206–213. [26] Y. S. K. Manjunath, L. Zhao, and X.-P. Zhang, “Segmented Learning for Metaverse Network Traffic Classification,” IEEE Internet of Things Journal, vol. 12, no. 19, pp. 40 068–40 082, 2025. [27] P. Pernot, “Stratification of Uncertainties Recalibrated by Isotonic Regression and Its Impact on Calibration Error Statistics,” arXiv preprint arXiv:2306.05180, 2023. [28] R. S. Kennedy, N. E. Lane, K. S. Berbaum, and M. G. Lilienthal, “Simulator Sickness Questionnaire: An Enhanced Method for Quantifying Simulator Sickness,” The International Journal of Aviation Psychology, vol. 3, no. 3, pp. 203–220, 1993. [29] J.-P. Stauffert, F. Niebling, and M. E. Latoschik, “CSQ-VR: A Novel Questionnaire for Measuring Cybersickness in Virtual Reality,” Frontiers in Virtual Reality, vol. 4, p. 1130864, 2023. [30] H. K. Kim, J. Park, Y. Choi, and M. Choe, “Virtual Reality Sickness Questionnaire (VRSQ): Motion Sickness Measurement Index in a Virtual Reality Environment,” Applied Ergonomics, vol. 69, pp. 66–73, 2018. [31] Y. S. K. Manjunath, J. Gao, and L. Zhao, “Extended Reality Network Traffic & Quality of Experience,” 2025. [32] C. R. Harris et al., “Array Programming with NumPy,” Nature, vol. 585, no. 7825, pp. 357–362, Sep. 2020. [33] Wes McKinney, “Data Structures for Statistical Computing in Python,” in Proceedings of the 9th Python in Science Conference, Stéfan van der Walt and Jarrod Millman, Eds., 2010, pp. 56–61. [34] L. Buitinck et al., “API Design for Machine Learning Software: Experiences from the scikit-learn Project,” in ECML PKDD Workshop: Languages for Data Mining and Machine Learning, 2013, pp. 108–122. [35] M. Abadi et al., “TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems,” 2015, software available from tensorflow.org. [Online]. Available: https://www.tensorflow.org/ [36] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., “PyTorch: An Imperative Style, High-Performance Deep Learning Library,” Advances in Neural Information Processing Systems, vol. 32, 2019. [37] N. Posocco and A. Bonnefoy, “Estimating Expected Calibration Errors,” in International Conference on Artificial Neural Networks. Springer, 2021, pp. 139–150. [38] A. A. De Hond, E. W. Steyerberg, and B. Van Calster, “Interpreting Area Under the Receiver Operating Characteristic Curve,” The Lancet Digital Health, vol. 4, no. 12, pp. e853–e855, 2022. [39] J. F. Shortle, J. M. Thompson, D. Gross, and C. M. Harris, Fundamentals of Queueing Theory. John Wiley & Sons, 2018.

15

Yoga Suhas Kuruba Manjunath (Member, IEEE) received the Bachelor of Engineering degree in electronics and communication from Visvesvaraya Technological University, Belagavi, India, in 2014, the Master of Engineering degree in artificial intelligence from Toronto Metropolitan University (TMU) (formerly Ryerson University), Toronto, ON, Canada, in 2021, and the Ph.D. degree from the Department of Electrical and Computer Engineering, TMU, in 2025. He has over five years of industry experience as an Internet-of-Things (IoT) architect, developing IoT stacks for the dairy and hospitality industries that have impacted over two hundred thousand customers. His combined hardware and software expertise has contributed to numerous projects in these sectors. Upon returning to academia, he focused on advancing his research skills, resulting in publications in leading conferences and journals, including IEEE GLOBECOM, IEEE WF-IoT, and Electronic Commerce Research and Applications. He actively participates in communications-related projects at the Communications and Signal Processing Applications Laboratory and Ubiquitous Intelligent Communication and Computing at TMU. His current research interests include AI-based IoT solutions and virtual-reality network optimization for quality of service. Dr. Manjunath received the Best Team Award at the IEEE Leaders of Tomorrow event organized by IEEE Toronto. He serves as a peer reviewer for several journals, including IEEE Transactions on Wireless Communications, IEEE Transactions on Network and Service Management, IEEE Internet of Things Journal, IEEE Open Journal of the Communications Society, Electronic Commerce Research and Applications, and The Journal of Supercomputing. He serves as vice chair of the IEEE Vehicular Technology Chapter of the IEEE Toronto Section and is a member of the IEEE Vehicular Technology Society and the IEEE Communications Society.

Jie Gao (Senior Member, IEEE) is an Assistant Professor with the School of Information Technology, Carleton University, Ottawa, ON, Canada. His research interests include machine learning for communications and networking, XR and immersive communications, and emerging network technologies for 6G. He is serving or has served as an Editor for IEEE Transactions on Cognitive Communications and Networking, IEEE Open Journal of the Communications Society, and the Vehicular Technology Section of IEEE Access. He has cochaired symposia, tracks, and workshops at IEEE conferences, including IEEE GLOBECOM, VTC, ICCC, and INFOCOM. He received the IEEE Vehicular Technology Society Open Journal of Vehicular Technology Best Paper Award in 2025, the IEEE Best Land Transportation Paper Award in 2024, and the Wisconsin Space Grant Consortium Early-Stage Investigator Grant Award in 2021.

Lian Zhao (Fellow, IEEE) received the Ph.D. degree from the Department of Electrical and Computer Engineering (ELCE), University of Waterloo, Canada, in 2002. She joined the Department of Electrical and Computer Engineering at Toronto Metropolitan University (formerly Ryerson University), Canada, in 2003. Her research interests are in the areas of wireless communications, resource management, mobile edge computing, IoT/IoV networks, and machine learning for communications. She has been an IEEE Communication Society (ComSoc) and IEEE Vehicular Technology (VTS) Distinguished Lecturer (DL); received the Best Land Transportation Paper Award from IEEE Vehicular Technology Society in 2016 and 2024, Best Paper Award from the 2013 International Conference on Wireless Communications and Signal Processing (WCSP), and the Canada Foundation for Innovation (CFI) New Opportunity Research Award in 2005. She has been serving as an Editor for IEEE Transactions on Wireless Communications, IEEE Internet of Things Journal, and IEEE Transactions on Vehicular Technology (2013-2021). She serves as a TPC Chair for VTC2025Fall, a co-Chair of Wireless Communication Symposium for IEEE Globecom 2020/2025, and IEEE ICC 2018; Finance co-Chair for 2021 ICASSP; Local Arrangement co-Chair for IEEE VTC Fall 2017 and IEEE Infocom 2014. She has been an elected member for the Board of Governor (BoG) of VTS since 2023.