TG-DIN: Theory-Guided Demand Inference Network for Generalizable QoS Measurement and Prediction Fuliang Yang and Feng Ye
arXiv:2605.15550v1 [cs.NI] 15 May 2026
[email protected],[email protected] University of Wisconsin-Madison Madison, Wisconsin, USA
Abstract
1
In this paper, we introduce TG-DIN, a theory-guided demand inference network that infers latent user demand from observable network quality-of-service (QoS) measurements. Rather than directly predicting QoS outcomes using black-box models, TG-DIN explicitly models latent demand as an intermediate variable and links it to observable behavior through a differentiable theory layer grounded in scheduling and queuing principles. This design yields an interpretable, mechanism-consistent representation of user demand that is directly applicable to downstream tasks such as congestion diagnosis, resource allocation, capacity planning, and policy evaluation. The theory layer further enables a principled randomized training regime that exposes the model to diverse yet physically meaningful operating conditions without requiring labeled demand data. Extensive synthetic experiments show that TG-DIN generalizes robustly across capacities, demand levels, and traffic patterns, substantially outperforming purely data-driven baselines under distribution shift. Moreover, when trained exclusively on synthetic data and applied directly to real packet traces, TG-DIN accurately recovers per-user allocation structure in shared-link scenarios. Together, these results demonstrate the effectiveness of theory-guided inductive biases for achieving transferable, deployment-ready inference in dynamic network environments.
Network users and operators can readily observe quality-of-service (QoS) metrics such as throughput, delay, and packet loss. However, these metrics reflect only the network’s final outcomes and conceal the underlying causes that generate them. Identical QoS observations may arise from fundamentally different conditions. For example, low throughput may be caused by low user demand, limited link capacity, contention with other flows, or unfavorable scheduling. Distinguishing among these causes is critical for effective network management and diagnosis. As a result, latent user demand is a key quantity for explaining QoS behavior and enabling downstream tasks such as congestion diagnosis, resource allocation, capacity planning, and scheduling-policy evaluation [20]. Despite its importance, actual latent demand is typically bypassed in existing data-driven approaches that usually model observable traffic traces or QoS measurements directly [23]. While convenient, this black-box formulation often struggles in real deployments where network conditions vary widely and rely heavily on representative training data from the target environment. Their performance degrades when deployment conditions differ from those seen during training [23, 26]. When operating regimes shift or novel conditions arise, such models typically require new data collection and retraining, which can be costly and impractical in operational settings. In this work, we propose TG-DIN, a theory-guided demand inference network for inferring latent user demand for generalizable QoS measurement and prediction. The key idea is to model latent user demand as an explicit intermediate variable. We then relate this variable to observable QoS metrics through a differentiable theory layer based on known network mechanisms, including scheduling and queuing dynamics. This design separates user side demand from network side service limitations. As a result, the inferred demand is more interpretable and less dependent on spurious correlations in the data. Crucially, the theory layer enables a principled randomized data generation regime that exposes the model to diverse yet physically meaningful combinations of capacity, demand, traffic patterns, and change modes. This allows the model to be trained without requiring direct labels for latent demand and eliminates the need to collect environment-specific datasets for each new operating regime. The resulting model is trained once and applied directly across a wide range of conditions, which improves cross-regime generalization and reduces dependence on target-environment calibration. We evaluate TG-DIN against purely data-driven baselines that directly predict observable traffic outcomes. Extensive experiments under both matched and shifted regimes demonstrate that our approach achieves superior robustness on synthetic benchmarks and transfers effectively to
CCS Concepts • Networks → Network measurement.
Keywords Network measurement, generalization, neural network ACM Reference Format: Fuliang Yang and Feng Ye. 2026. TG-DIN: Theory-Guided Demand Inference Network for Generalizable QoS Measurement and Prediction. In Proceedings of Make sure to enter the correct conference title from your rights confirmation email (Conference acronym ’XX). ACM, New York, NY, USA, 10 pages. https: //doi.org/10.1145/nnnnnnn.nnnnnnn
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. Conference acronym ’XX, Woodstock, NY © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/2018/06 https://doi.org/10.1145/nnnnnnn.nnnnnnn
Introduction
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
real-traffic traces spanning diverse capacities, demand levels, and traffic patterns. Our main contributions are summarized as follows: (1) We propose TG-DIN, a theory-guided demand inference network that recovers latent user demand for generalizable QoS measurement and prediction. the framework provides interpretable and mechanism-consistent inference by introducing latent demand as an explicit intermediate variable and connecting it to QoS through a theory layer. (2) We introduce a randomized data generation regime enabled by the theory layer, producing diverse yet physically plausible scenarios across capacities, traffic patterns, demand levels, and drift types without requiring labeled demand data. (3) We demonstrate through comprehensive synthetic and real traffic evaluations that TG-DIN achieves strong cross-regime robustness and eliminates the need to recollect representative traces or fine-tune models for each deployment condition.
2 Related work 2.1 Network Measurement and QoS Prediction Prior work has extensively studied the measurement, inference, and prediction of observable network performance metrics such as latency, throughput, and end-to-end path behavior. Early systems such as iPlane [18] and iPlaneNano [19] demonstrated that largescale path properties can be inferred from distributed measurements and lightweight probes. These studies established performance estimation as a central theme in network measurement research. A parallel line of work modeled and synthesized Internet delay spaces, showing that structural properties of network performance can be captured using compact statistical representations [31]. Building on this foundation, more recent studies have advanced latency distribution and tail latency estimation in modern settings including network functions, datacenter fabrics, and large-scale service deployments [12, 32, 34]. A complementary body of work focuses on learning direct mappings from measured network signals to application level performance, such as quality-of-experience (QoE) prediction for video streaming and QoS/QoE characterization for real-time applications [4, 6]. While these approaches are highly effective for monitoring, diagnosis, and forecasting, they treat QoS or QoE as the final prediction target rather than as indirect observations of underlying user demand. In contrast, our work focuses on latent demand inference in a shared-link setting, where each user’s request rate is coupled to a common bottleneck through a scheduling policy and is therefore only observable via its resulting throughput and QoS. Recovering user demand from such measurements requires explicitly disentangling user-side demand from network-side service constraints. This is a problem that lies outside the scope of existing network measurement and QoS-prediction literature. Our approach addresses this gap directly by introducing latent demand as an explicit intermediate variable and grounding its relationship to observable behavior in network theory.
Yang et al.
2.2
Traffic Engineering under Demand Uncertainty
A related line of work focuses on the control side of network management. Given historical observations, these systems directly produce routing or allocation decisions rather than explicitly recovering hidden per user demand. Classic wide area network systems such as B4 [13] and SWAN [10] showed that traffic engineering must cope with continually varying traffic patterns and service objectives. They formulated network optimization as a centralized allocation problem driven by periodically measured demand. These systems rely on classical optimization pipelines in which demand estimation and control are largely decoupled. More recent work has increasingly embraced learning-based approaches to bypass or compress this pipeline. DOTE [23] argues for training traffic engineering decision models directly on historical demand data rather than predicting future demand as an explicit intermediate step. TEAL [29] and subsequent neural WAN frameworks [1] further illustrate how learning can accelerate or generalize traffic engineering decisions across changing network conditions. Real time systems such as RedTE [9] also target sub second burst driven control loops that operate close to the physical link. While these systems operate different scales, they share a key methodological choice: observable history is mapped directly to future targets or control actions without introducing latent demand as an explicit intermediate variable. In contrast, our work makes latent demand the central object of inference and constrains its relationship to observable allocations through a differentiable scheduling and queuing layer grounded in network theory. This distinction allows our approach to separate user-side demand from network-side service constraints, rather than collapsing them into a single control mapping.
2.3
Hidden Traffic Demand Estimation
Prior work has long recognized the importance of recovering hidden traffic demand from indirect network measurements. The most closely related line of research is traffic-matrix estimation that aims to infer source–destination demand volumes from observations such as link loads, routing configurations, or partial flow statistics [20, 33]. More recent studies have extended this direction by applying learning-based estimators and generative models to traffic-matrix recovery, enabling more flexible inference from incomplete or noisy measurements [24, 28, 30]. These efforts share a core motivation with our work. Observable network measurements alone are often insufficient for effective network management, so hidden demand must be inferred rather than directly measured. However, existing demand estimation work typically operates at a much coarser granularity than the setting we consider. In particular, traffic-matrix estimation focuses on aggregate source–destination demand in backbone or interdomain networks, whereas our work targets per-user latent demand in a shared-link environment. The observation models also differ fundamentally. Traditional approaches infer global demand from network-wide measurements and routing constraints, while we infer user-level demand from local QoS observations at a bottleneck link. Our objective is therefore not to reconstruct a network-wide traffic matrix, but to identify the hidden user-side demand state that explains the observed per-user
TG-DIN: Theory-Guided Demand Inference Network for Generalizable QoS Measurement and Prediction
throughput and QoS behavior. This shift in granularity and observation model places our work outside the scope of conventional traffic-matrix estimation and motivates a distinct approach based on theory-guided inference at the level of individual users and shared bottlenecks.
2.4
Deep Learning in Network Management
Deep learning has seen growing adoption in network measurement, traffic engineering, and QoS management. A dominant paradigm in this line of work is to learn direct mappings from historical observations to future traffic volumes, performance metrics, or control decisions, without explicitly modeling user demand as a latent variable. In network performance prediction, representative examples include PERCEIVE, which employs a two-stage LSTM architecture for short-horizon cellular uplink throughput prediction from real-time scheduling patterns [16]; ensemble GRU-LSTM models for traffic prediction in research and education networks [25]; and transformer-based models such as Temporal Fusion Transformer for mobile-network traffic forecasting [15]. Recent survey work similarly identifies recurrent and transformer-based architectures as the dominant deep-learning families in network traffic prediction [3]. Beyond raw traffic forecasting, deep learning has also been applied to user- and service-level performance prediction. Recent efforts include spatial-context-aware QoS forecasting [35] and graph-attention–based collaborative learning for temporal QoS prediction [11]. Together, these works illustrate a common methodological pattern: learned sequence models map historical measurements directly to future observable outcomes, whether the prediction target is traffic volume, throughput, or QoS. A related practical challenge for such models is distribution shift after deployment. Prior work has shown that concept drift arises in operational networks, for example in cellular-network prediction settings [17]. However, most existing concept-drift studies in networking focus on the security domain, such as intrusion detection, malware detection, or traffic classification—rather than continuous performance prediction [2, 14, 27]. As a result, these techniques are not directly applicable to QoS inference under capacity and demand shifts. In contrast to these direct-prediction and adaptation-based approaches, the key distinction of our method lies not in the use of deep sequence models themselves, but in what the models are trained to infer. Rather than directly predicting observable outcomes, we explicitly recover latent user demand and constrain its relationship to per-user allocation and QoS through a structured, differentiable theory layer. This design enables interpretable inference, reduces reliance on environment-specific correlations, and improves robustness under distribution shift.
observable network behavior. Formally, let 𝑑𝑡 ∈ R𝑈+ denote the latent per-user demand at time 𝑡, where 𝑈 is the number of users. We seek a predictor 𝑑ˆ𝑡 = 𝑔𝜙 (𝑥𝑡 −𝐾:𝑡 ), (1) where 𝑔𝜙 is a neural demand inference model. The inferred demand is then passed through a theory-guided forward model 𝑦ˆ𝑡 = 𝑓theory (𝑑ˆ𝑡 , 𝐶𝑡 , 𝜋), (2) where 𝐶𝑡 is the link capacity, 𝜋 denotes the scheduling policy, and 𝑦ˆ𝑡 represents the predicted observable network outcomes. In our setting, these outcomes include throughput and additionally include other QoS variables such as delay and loss. This formulation turns QoS analysis into a latent demand inference problem constrained by network theory. Instead of directly fitting a black box mapping from observed inputs to observed outputs, we require the inferred hidden state to produce observable behavior through network theory.
3.2
Latent Demand Inference
The first component of our framework is a latent demand predictor. Given the recent observation history 𝑥𝑡 −𝐾:𝑡 , the predictor outputs a nonnegative demand estimate for each user: 𝑑ˆ𝑡 = 𝑔𝜙 (𝑥𝑡 −𝐾:𝑡 ), 𝑑ˆ𝑡 ≥ 0. (3) This intermediate representation plays a central role in our design. Directly predicting observable traffic outcomes from QoS features can capture correlations present in the training data, but it does not explicitly distinguish between what the user demands and what the network delivers. By introducing latent demand as an explicit intermediate state, our framework models the hidden cause underlying observed QoS behavior rather than collapsing demand and service effects into a single prediction. Without loss of generality, we implement 𝑔𝜙 as a neural predictor operating on a window of recent observations. The input feature vector is processed by a multi layer perceptron (MLP) consisting of three fully connected hidden layers with ReLU activations, followed by a two-dimensional output layer and a Softplus nonlinearity to enforce non-negativity. Although more sophisticated architectures could be employed, our experimental results show that this simple design already achieves strong accuracy and robust generalization across a wide range of capacity and demand conditions. An overview of the complete model architecture is shown in Fig. 1. Theory-Guided Latent-Demand Model neural demand inference → Theory layer → Observable traffics link capacity feeds allocator
2 1
At each time window 𝑡, we observe traffic traces through throughput and queue related QoS, together with the current link capacity information through actual measurement or estimation. Let 𝑥𝑡 −𝐾:𝑡 represent the causal observation history from the most recent 𝐾 windows. However, the underlying per-user traffic demand is latent and not directly measurable from these observations alone. Our goal is to infer this latent demand and use it to explain the resulting
Neural demand inference
Observable traffics
Allocator
latent per-user demand
allocated bandwidth
Norm
Softplus Head FC 128 + ReLU
Randomized Dataset
Theory layer per-user rate split
per-user traffic features
0
3
KEY BOTTLENECK
past observations link capacity
3 Theory-Guided Demand Inference Network 3.1 Problem Formulation
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Queue step
Buffer occupancy, time interval feed queue step
Predicted per-user traffic
Sample (C, d, pattern) Simulate 200 traces User-swap augmentation
Training objective (no demand supervision)
refresh every 30 epochs
Figure 1: Overall architecture of the proposed framework.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
3.3
Yang et al.
Theory Layer
The central contribution of TG-DIN is a differentiable theory layer that maps inferred latent demand to observable network outcomes. This layer explicitly encodes the network’s known forward mechanism and consists of two stages: scheduling and queuing. At each window 𝑡 of length Δ𝑡, each user 𝑖 has an observed buffer occupancy 𝐵 obs,𝑖,𝑡 , an inferred arrival-rate demand 𝑑ˆ𝑖,𝑡 from the head, and a shared link capacity 𝐶𝑡 . For readability, we omit the time index 𝑡 in the equations below. Scheduling module. The scheduling module computes a policydependent allocation from the inferred demand and available link capacity: 𝑎ˆ𝑡 = 𝑓sched (𝑑ˆ𝑡 , 𝐶𝑡 , 𝜋),
(4)
where 𝑎ˆ𝑡 denotes the per-user allocation predicted under bandwidth allocation policy 𝜋. Concretely, the scheduler first forms a stateaware effective demand that folds the current buffer into the arrival rate, 𝐵 obs,𝑖 + 𝑑ˆ𝑖 Δ𝑡 dem𝑖 = , (5) Δ𝑡 and then distributes the link capacity according to a work-conserving best-effort rule: Í if 𝑗 dem 𝑗 ≤ 𝐶, dem𝑖 , 𝑎ˆ𝑖 = (6) dem𝑖 𝐶 Í 𝑗 dem 𝑗 , otherwise, Í subject to 0 ≤ 𝑎ˆ𝑖 ≤ dem𝑖 and 𝑖 𝑎ˆ𝑖 ≤ 𝐶. Equation (6) is an instantiation of Generalized Processor Sharing [22], which is the theoretical basis of essentially all work-conserving fair-queueing disciplines. It is consistent with the default behavior of commodity IP routers [21][8], with the long-run fairness of the IEEE 802.11 DCF MAC used by Wi-Fi access points under saturation [5], and with the equilibrium behavior of TCP-style congestion control on a shared FIFO bottleneck [7]. The theory layer is, however, agnostic to the particular form of 𝑓sched : any differentiable rule satisfying the feasibility constraints above can be substituted for (6). Queuing module. The queuing module uses the inferred demand and allocation to produce observable QoS behavior: 𝑦ˆ𝑡 = 𝑓queue (𝑑ˆ𝑡 , 𝑎ˆ𝑡 , 𝑠𝑡 ),
(7)
where 𝑠𝑡 denotes the relevant queue state or history-dependent state variables. Concretely, within one window the arrival volume is 𝑑ˆ𝑖 Δ𝑡 and the serviceable volume is 𝑎ˆ𝑖 Δ𝑡, so the amount actually sent is capped by whichever is smaller, sent𝑖 = min 𝐵 obs,𝑖 + 𝑑ˆ𝑖 Δ𝑡, 𝑎ˆ𝑖 Δ𝑡 , (8) and the predicted throughput is 𝑟ˆ𝑖 = sent𝑖 /Δ𝑡 .
(9)
A queueing-delay proxy is obtained from the buffer divided by the allocated service rate (floored by 𝐴min and capped at 𝜏max ), 𝜏ˆ𝑖 = min 𝜏max,
𝐵 obs,𝑖 , max(𝑎ˆ𝑖 , 𝐴min )
(10)
and the loss rate is modeled as finite-buffer overflow with buffer size 𝐵 max and a clipping ceiling ℓmax , 𝑞𝑖rem = 𝐵 obs,𝑖 + 𝑑ˆ𝑖 Δ𝑡 − sent𝑖 , max(𝑞𝑖rem − 𝐵 max, 0) ℓˆ𝑖 = min ℓmax, . 𝐵 obs,𝑖 + 𝑑ˆ𝑖 Δ𝑡
(11) (12)
Eqs. (5)-(12) together define a closed form, parameter free forward ˆ that is differentiable almost everyˆ 𝜏, ˆ ℓ) map (𝑑,ˆ 𝐶, 𝐵 obs, Δ𝑡) ↦→ (𝑟, where, so losses defined on observable quantities can be back propagated through the theory layer to update the latent demand predictor. As a result, the model does not require direct labels for latent demand. Instead, it learns latent demand indirectly by finding hidden states that, when passed through the network mechanism, best explain the observed traffic. We emphasize that the supervision signal is the throughput 𝑟ˆ in equation (9), not the scheduler outˆ in a measured trace only throughput is observable, whereas put 𝑎: 𝑎ˆ is an internal scheduler variable with no direct counterpart in a packet capture. The latent demand 𝑑ˆ and the allocation 𝑎ˆ are therefore never directly supervised; they are inferred through backpropagation of the observable loss through Eqs. (9)-(12). The theory layer is critical for two reasons. First, it constrains the learned mapping to follow known scheduling and queuing behavior rather than arbitrary input-output correlations. Second, it provides the mechanism that enables our randomized training regime, described next.
3.4
Theory-Guided Randomized Regime Generation and Training Objective
Thanks to the newly developed theory layer, TG-DIN does not require pre-collecting or storing a large, fixed dataset. Instead, the model can continuously sample new link capacities, demand regimes, and traffic patterns during training, generate the corresponding synthetic traces on the fly, and optimize directly on these freshly instantiated scenarios. Because the mapping from latent demand to observable QoS is implemented as a differentiable forward theory layer, newly sampled regimes can be incorporated seamlessly into end-to-end training. During training, we explicitly randomize the primary axes along which real-world network conditions commonly vary: (1) link capacity, (2) temporal traffic pattern, (3) demand magnitude, and (4) demand change mode. Together, these factors define a broad family of network regimes over which the relationship between latent demand and observable QoS may differ substantially. This distinction is critical. A purely data-driven model trained on a static corpus may still overfit dataset-specific correlations, even when data augmentation or network simulator is applied. In contrast, our randomized regime generation is an integral part of the method itself. It systematically exposes the model to diverse yet physically meaningful operating conditions, while the theory layer ensures that each generated scenario obeys valid scheduling and queuing dynamics. In this sense, the theory layer enforces mechanism validity, while regime randomization provides coverage. As a result, the model is trained not merely to fit a fixed dataset, but to infer latent demand across an entire family of network regimes. This design is intended to improve robustness under capacity shifts, traffic-pattern changes, and demand-regime variation encountered in practice, while substantially reducing the
TG-DIN: Theory-Guided Demand Inference Network for Generalizable QoS Measurement and Prediction
need to recollect representative traces for each new deployment condition. A key property of our framework is that latent demand is never supervised with ground truth. Instead, the model is trained by matching the observable outputs produced by the theory layer to the observed network measurements. Concretely, the latent demand predictor first outputs 𝑑ˆ𝑡 , which is then passed through the differentiable scheduling and queueing modules to produce predicted observable quantities, including per-user throughput and additional QoS variables such as delay and loss. Our primary supervision is applied to per-user throughput, since throughput is the observable consequence of the interaction among demand, capacity, and the scheduling policy, and is the only per-user rate quantity that can be directly measured from a packet trace (whereas the scheduler’s internal allocation 𝑎ˆ𝑡 is not observable). Let 𝑟ˆ𝑡 and 𝑟𝑡 denote the predicted and observed per-user throughput, respectively. To improve stability across a wide dynamic range of traffic levels, we supervise throughput mainly in log-space: log
Lthr = MSE(log(1 + 𝑟ˆ𝑡 ), log(1 + 𝑟𝑡 )) .
(13)
This term prevents large throughput samples from dominating the objective and gives smaller throughput values a meaningful gradient signal. To retain calibration in the original scale, we additionally include a small linear-space throughput penalty: lin Lthr = SmoothL1(𝑟ˆ𝑡 , 𝑟𝑡 ).
(14)
The full throughput loss is log
lin Lthr = Lthr + 𝜆lin Lthr ,
(15)
where 𝜆lin is a small coefficient used to keep the log-space objective as the dominant training signal while still regularizing absolutescale errors. When auxiliary QoS observations are available, we also supervise the theory-layer predictions of those quantities. Let 𝑧ˆ𝑡 and 𝑧𝑡 denote the predicted and observed auxiliary QoS variables, respectively. We write the auxiliary loss as ∑︁ Laux = 𝜆𝑚 ℓ𝑚 (𝑧ˆ𝑡(𝑚) , 𝑧𝑡(𝑚) ), (16)
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
4 Evaluation 4.1 Experimental Setup Hardware and software platform. All experiments are conducted on a single Linux workstation (Ubuntu 22.04 LTS, kernel 6.8) equipped with an NVIDIA GeForce RTX 5090 GPU and 64 GB of system memory. Models are implemented in PyTorch 2.4 using Python 3.10, with NumPy 1.26, Pandas 2.2, and Matplotlib 3.8 for data processing and visualization. Randomized training scenarios. We instantiate the randomized training regime by regenerating 200 fresh synthetic traces at the start of each refresh round. Link capacities are sampled from the range [20, 600] Mbps and follow slowly varying trajectories over time. Per-user demand magnitudes are sampled from [1, 80] Mbps across four overlapping regime labels: small (up to 5 Mbps), light (approximately 1-8 Mbps), medium (approximately 3-20 Mbps), and heavy (10 Mbps and above, up to the configured maximum). Per-user traffic patterns are independently sampled from the set continuous, on-off. The continuous pattern corresponds to approximately steady traffic over time, while the on-off pattern represents bursty behavior with alternating active and idle periods. A new training round is refreshed every 30 epochs; as a result, the model is exposed over the course of training to a large and diverse set of (capacity, demand, pattern) configurations without relying on a fixed trace corpus. TG-DIN training. The proposed model is trained end-to-end using an observable reconstruction loss, without access to truth demand labels. We use the Adam optimizer with a learning rate of 10−4 , weight decay of 10−5 , and a batch size of 256. Within each refresh round, synthetic traces are split into 80%/10%/10% train, validation, and test partitions at the trace level. To discourage reliance on user identity, we apply a user-swap data augmentation with probability 0.5 per batch, in which the observable features and supervision targets of the two users are exchanged simultaneously. Model checkpointing and early stopping are not driven by the per-round validation split, but instead by performance on a fixed, stratified global calibration and evaluation set constructed at training start. This set spans 8 capacity bands × 4 demand regimes × 4 traffic-pattern combinations, ensuring consistent coverage of operating conditions throughout training. A summary of the training protocol and key hyperparameters is provided in Table 1.
𝑚∈ M
where M indexes the supervised QoS variables and ℓ𝑚 (·, ·) denotes the corresponding per-variable loss. In our setting, these auxiliary terms can include delay-related and loss-related supervision, depending on the experiment configuration. The overall training objective is therefore L = Lthr + Laux .
(17)
This objective couples the neural predictor and the theory layer tightly. The predictor proposes latent demand, the theory layer maps it to observable network behavior, and the loss rewards latent states whose induced observable outcomes best match the data. Because gradients propagate through the differentiable scheduling and queuing theory, the model can learn latent demand without requiring direct demand labels.
4.2
Baselines and Benchmarking
We compare against purely data-driven direct predictors that operate on the same causal observation window as our method, but map observable features directly to per-user throughput outputs without inferring latent demand or incorporating an explicit theory layer. All baselines use the same window length (𝐾=5), identical feature construction, and the same throughput loss function as the proposed model, ensuring a fair comparison. • Direct-GRU-LSTM [25]. Adapted from the GRU-LSTM traffic forecaster, this baseline represents the recurrent directprediction family. The model consists of a stacked GRU and LSTM followed by a two-layer MLP prediction head, with a final softplus activation to produce non-negative per-user allocation estimates.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 1: Training setup and key hyperparameter for the proposed model and baselines. Category
Randomized scenarios
TG-DIN
Baselines
Parameter
Yang et al.
Table 2: Scenario definitions used in the synthetic test grid. All rates in Mbps.
Value
Scenario
Observation window 𝐾
5 windows (0.2 s per window)
Traces per refresh round
200
Refresh frequency
every 30 epochs
Link capacity range
[20, 600] Mbps
demandOnly / bursty, 0-12, 𝜇 ≈ 2.6 → cont., 1.2-5.0, 𝜇 ≈ 3.0 𝑢 0 level 2.9 ch. 𝑢 0 (unch.) demandOnly / bursty YT-like, 0-75, 𝜇 ≈ cont., 6.5-14, 𝜇 ≈ 10 𝑢 1 level ch. 𝑢 1 17-19 (unch.) (shift) patternOnly / high-rate, 37-62, 𝜇 ≈ 50 cont., 1.1-5.1, 𝜇 ≈ 3.0 𝑢 0 pattern ch. 𝑢 0 (unch.) (unch.) patternOnly / bursty YT-like, 0-75, 𝜇 ≈ sparse, 0-2.5, 𝜇 ≈ 1.3 𝑢 1 duty ch. 𝑢 1 (unch.) 17-19 (unch.) patternDemand / cont., 5.6-14, 𝜇 ≈ 10 cont., 1.1-5.1, 𝜇 ≈ 3.0 𝑢 0 joint ch. 𝑢 0 (unch.) patternDemand / bursty, 0-75, 𝜇 ≈ 17-19 bursty, 0-12, 𝜇 ≈ 6.4-6.7 𝑢 1 joint ch. 𝑢 1 (unch.)
Per-user demand range
[1, 80] Mbps
Demand regimes
small ( ≤ 5), light (1-8), medium (3-20), heavy ( ≥ 10) Mbps
Traffic patterns Optimizer
continuous (steady), on-off (bursty) Adam
Learning rate
10 −4
Weight decay
10 −5
Batch size
256
Optimizer
Adam
Learning rate
10 −3
LR scheduler
ReduceLROnPlateau ( × 0.5, patience 3)
Batch size
256
Gradient clipping (ℓ2 )
5.0
Maximum epochs
60
Early stopping patience
10 epochs
Train / validation split
90% / 10% (single-capacity data)
• Direct-TFT-style [15]. Adapted from Temporal Fusion Transformer–based traffic prediction models, this baseline represents the attention-based direct-prediction family. Our implementation includes a linear input projection, an LSTM encoder, a multi-head self-attention block, and a feed-forward prediction head with softplus output. We refer to this model as TFT-style because it retains the core sequence modeling and attention mechanisms, while omitting components that are not required in our single-step, two-user setting. Baseline training. All direct-prediction baselines share the same input features, log-based normalization, batch size (256), and throughput loss formulation (log-space MSE + 0.01· smooth-𝐿1 ) as the proposed method. Delay and loss-rate terms are excluded, as the baselines output throughput only. Training uses the Adam optimizer with a learning rate of 10−3 , ReduceLROnPlateau scheduling (factor 0.5, patience 3), gradient clipping with ℓ2 -norm 5.0, and a maximum of 60 epochs with early stopping (patience 10). Ten percent of training traces are held out for validation. To reflect realistic deployment and data-collection constraints, each baseline is trained on a single fixed-capacity synthetic trace corresponding to one operating condition, and then evaluated under shifted capacities and drift scenarios. This setup mirrors common direct-prediction workflows, in which representative data are available only for the current environment and not for all operating regimes that may be encountered after deployment.
𝑢 0 profile
𝑢 1 profile
Change
Target-adapted fine-tuning baseline. To model a standard adaptation strategy under concept drift, we additionally fine-tune both direct-prediction baselines using small calibration sets drawn from the target capacity regime. Such fine-tuning is a widely used approach for updating data-driven predictors when the deployment distribution shifts. Although the calibration sets are synthetic in our experiments, they serve as proxies for newly collected target-environment traces in practice. We consider calibration budgets of 1% and 5% of target windows, together with two adaptation strategies: full fine-tuning of all model parameters and last-layer fine-tuning of the output head only. This gives the direct baselines explicit access to target-condition data, whereas the proposed theory-guided model is evaluated without any fine-tuning. Synthetic test grid. Our primary synthetic benchmark spans seven test capacities, 20, 40, 60, 120, 200, 280, 360 Mbps, crossed with six drift scenarios per capacity. The six scenarios are obtained by combining three scenario families, i.e. demandOnly, patternOnly, and patternDemand, with two change indicators changeU0 and changeU1. Each synthetic test trace contains a single concept-drift event introduced at the midpoint of the sequence. We adopt a two-user setting in which 𝑢 0 represents a higher-rate, bursty on-off flow, while 𝑢 1 represents a lower-rate, smoother continuous flow. The three scenario families capture distinct types of drift and are defined as follows: • demandOnly: only the average demand level of one user changes, while the temporal traffic pattern remains fixed; • patternOnly: the temporal traffic pattern changes, while the mean demand level remains approximately unchanged; • patternDemand: both the demand level and the temporal traffic pattern change simultaneously. Representative traffic profiles and rate ranges for these scenarios are summarized in Table 2. Using this test grid, we evaluate synthetic transfer as follows: (i) pooled and per-scenario cross-capacity generalization without adaptation; and (ii) target-adapted fine-tuning under capacity and demand shifts.
4.3
Pooled and Per-Scenario Cross-Capacity Generalization without Adaptation
We first evaluate cross-capacity generalization by comparing the proposed theory-guided model against six single-capacity directprediction baselines on the synthetic test grid. Each baseline is
TG-DIN: Theory-Guided Demand Inference Network for Generalizable QoS Measurement and Prediction
(a) Mean RMSE
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
(b) Mean relative MAE
(c) Mean Pearson 𝑟
Figure 2: Pooled cross-capacity synthetic transfer results without adaptation.
(a) demandOnly, 𝐶 = 20
(b) demandOnly, 𝐶 = 60
(c) demandOnly, 𝐶 = 280
(d) demandOnly, 𝐶 = 360
(e) patternDemand, 𝐶 = 20
(f) patternDemand, 𝐶 = 60
(g) patternDemand, 𝐶 = 280
(h) patternDemand, 𝐶 = 360
Figure 3: Representative synthetic time-series comparisons for two scenario families (demandOnly and patternDemand) across four test capacities. trained at a single source capacity (20, 40, or 60 Mbps) and evaluated across all target capacities without any form of target-side adaptation. Fig. 2 summarizes the resulting pooled RMSE, relative MAE, and Pearson-𝑟 trends across all testing capacities. As detailed in Table 3, the direct-prediction baselines exhibit strongly capacity-specific behavior: performance is best near the training capacity and degrades rapidly under capacity shift. For instance, the GRU-LSTM baseline trained at 20 Mbps achieves an RMSE of 2.78 Mbps at its source condition but deteriorates to 13.47 Mbps Table 3: Cross-capacity RMSE (Mbps) on the synthetic sweep. Test capacity (Mbps) Method
20
40
60
GRU-LSTM (20) GRU-LSTM (40) GRU-LSTM (60) TFT-style (20) TFT-style (40) TFT-style (60)
2.78 7.74 11.81 2.50 4.01 10.95
7.06 3.30 7.65 7.25 3.28 7.07
11.92 13.51 13.49 13.48 13.47 7.42 9.54 9.51 9.49 9.48 4.60 5.41 5.39 5.38 5.38 12.63 14.47 14.47 14.48 14.48 7.18 8.99 9.02 9.04 9.06 3.53 5.28 5.46 5.60 5.71
TG-DIN
2.81
1.89
3.11
120
4.08
200
1.50
280
0.94
360
1.81
at 360 Mbps. A similar pattern is observed for the TFT-style baseline trained at 20 Mbps, whose RMSE increases from 2.50 Mbps to 14.48 Mbps over the same range. These results highlight the narrow specialization of single-source direct predictors and their limited ability to extrapolate beyond the training regime. In contrast, the theory-guided model maintains stable performance across the entire capacity range. Its RMSE remains bounded by 4.08 Mbps across all test capacities and attains its lowest errors at intermediate, unseen capacities (0.94 Mbps at 280 Mbps and 1.50 Mbps at 200 Mbps). This behavior suggests that the randomized theory-guided training regime enables substantially more transferable representations than direct predictors trained at a single operating point. The additional pooled metrics reinforce this conclusion. Relative MAE demonstrates that the observed gains are not driven solely by scaling effects at high capacities, while Pearson-𝑟 indicates that the TG-DIN preserves temporal allocation trends more faithfully across capacities. Together, these metrics show that the proposed approach captures both the magnitude and structure of per-user allocation dynamics under capacity shift. To further illustrate these differences, Fig. 3 visualizes representative held-out allocation trajectories from two scenario families (demandOnly and patternDemand) at four test capacities (20, 60,
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
(a) RMSE.
Yang et al.
(b) Relative MAE.
(c) Pearson 𝑟 .
Figure 4: Target-adapted fine-tuning under concept drift. Direct baselines are adapted using 1% or 5% target-capacity calibration data with either full-model or last-layer updates. 280, and 360 Mbps). Across all eight examples, the theory-guided model produces throughput that remain visibly closer to the ground truth than those of the two 60-Mbps direct baselines for both users. In particular, the direct baselines tend either to overshoot the dominant 𝑢 0 bursts or to collapse toward nearly constant predictions for the smaller 𝑢 1 flow. By contrast, the theory-guided model more consistently preserves both allocation levels and temporal structure. At lower capacities, the theory-guided model shows a mild tendency to over-predict the dominant flow under congestion. However, this bias is modest relative to the pronounced structural mismatches observed in the direct baselines. Importantly, the proposed model continues to track the correct switching behavior and relative user dynamics over time, aligning closely with the aggregate cross-capacity trends discussed above.
4.4
Target-Adapted Fine-Tuning Under Capacity and Demand Shifts
We next examine whether limited target-side calibration data can mitigate the cross-capacity generalization gap of direct-prediction baselines. To emulate a standard concept-drift adaptation workflow, we fine-tune both baseline families using small calibration sets drawn from the shifted target regime. All baselines are initialized from the 60-Mbps source checkpoint and adapted to target capacities 20, 40, 120, 200, 280, 360 Mbps using calibration budgets of 1% and 5% of target windows. For each target capacity, the calibration data are formed by aggregating small subsets from all six drift scenarios, ensuring exposure to both capacity and demand-level variation. We consider two adaptation strategies: full fine-tuning of all model parameters and last-layer fine-tuning of the output head only. Fig. 4 summarizes the resulting RMSE and Pearson-𝑟 trends. As expected, target-side fine-tuning substantially improves the performance of direct predictors, with the largest gains achieved under the 5% full-update setting. These results confirm that even small amounts of calibration data can be effective in correcting capacity-specific mismatch in purely data-driven models. However, the benefits of fine-tuning are largely localized to the calibrated conditions and do not consistently restore strong performance across all testing capacities. Table 4 quantifies these effects. At lower capacities, fine-tuning enables dramatic error reduction: for example, the GRU-LSTM baseline improves from 11.81 to 2.66 Mbps RMSE at 20 Mbps, while the
Table 4: Target-adapted fine-tuning under concept drift. RMSE (Mbps) of the 𝐶=60 Mbps-trained source baselines with 1%/5% calibration budgets
Model
GRULSTM (synth, 𝐶=60)
TFT-style (𝐶=60)
TG-DIN
Test capacity (Mbps)
Method 20
40
120
200
280
360
No-adapt +1% last +1% full +5% last +5% full No-adapt +1% last +1% full +5% last +5% full
11.81 8.38 3.19 3.37 2.66 10.95 7.74 2.93 3.36 1.90
7.65 4.78 4.54 3.17 3.08 7.07 4.43 3.74 2.57 2.53
5.41 5.27 5.14 4.96 4.82 5.27 4.68 4.99 3.80 3.44
5.39 5.26 5.14 4.96 4.82 5.46 4.71 5.01 3.82 3.48
5.38 5.25 5.14 4.96 4.82 5.60 4.74 5.02 3.84 3.49
5.38 5.25 5.14 4.95 4.82 5.71 4.76 5.03 3.85 3.50
No-adapt
2.80
1.89
4.07
1.50
0.94
1.81
TFT-style baseline improves from 10.95 to 1.90 Mbps under the same setting. These results highlight the effectiveness of calibrated adaptation when representative target data are available. Nonetheless, at higher capacities the adapted baselines continue to lag behind the theory-guided model. Even with 5% full fine-tuning, the strongest TFT-style baseline attains RMSEs of 3.48, 3.49, and 3.50 Mbps at 200, 280, and 360 Mbps, respectively—substantially higher than the corresponding errors of 1.50, 0.94, and 1.81 Mbps achieved by the theory-guided model without any target-capacity adaptation. The table further shows that full-model fine-tuning consistently outperforms last-layer-only updates, indicating that capacity and demand shifts cannot be fully addressed by simple output-level recalibration. Instead, they require deeper representational adjustment within the model. In contrast, the theory-guided approach maintains robust performance across capacities without access to any target-side data, underscoring its reduced dependence on post-deployment recalibration. Overall, these results indicate that while small-budget target adaptation can improve direct predictors locally, the theory-guided model achieves superior cross-capacity robustness without any fine-tuning.
4.5
Real-Traffic Transfer
To evaluate whether the proposed theory-guided demand representation transfers beyond the synthetic setting, we conduct a
TG-DIN: Theory-Guided Demand Inference Network for Generalizable QoS Measurement and Prediction
(a) 𝐶 = 20 Mbps
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
(b) 𝐶 = 60 Mbps
(c) 𝐶 = 120 Mbps
Figure 5: Representative real-traffic allocation traces under controlled bottleneck capacities. preliminary real-traffic experiment in a controlled two-user network environment. Our goal is to assess whether the learned latent demand representation and the theory-layer structure remain meaningful when the observable QoS signals are derived from real packet traces rather than from simulated traffic. In particular, we examine whether the model can still infer plausible allocation structure when exposed to real application behavior and measurement noise. Our real-traffic testbed consists of a Linux machine acting as a shared network gateway that provides Internet access to two client laptops over Wi-Fi. We adopt this design to leverage both high-capacity Internet connectivity and a high-rate wireless link, while retaining sufficient flexibility and control to emulate a range of access-network conditions experienced by end users. Both laptops therefore connect to the same wireless network while generating traffic independently, allowing them to share a common downstream bottleneck. Each laptop emulates a distinct user in the shared-link scenario. One device (user 0) generates YouTube video streaming traffic, while the other (user 1) runs a real-time video conferencing application (Microsoft Teams). On the Linux gateway, we use the Linux traffic-control framework (tc) to enforce a configurable bottleneck rate on the outbound path, ensuring that both Wi-Fi flows traverse and compete over the same explicitly rate-limited link. This setup creates controlled contention between the two users and induces observable interactions in their achieved throughput. Without loss of generality, we collect traces under three bottleneck capacities: 20, 60, and 120 Mbps. The bottleneck rate is explicitly configured at the gateway for each experiment, so the underlying capacity is known by construction. During each run, packet-level traces are recorded at the gateway and subsequently post-processed using the same fixed-window aggregation pipeline employed in the synthetic experiments. Packets are aggregated into time windows and converted into traffic representations with the observable QoS features required by the model, ensuring that synthetic and real-traffic evaluations share an identical input interface. This experimental setup enables a direct test of whether the theory-guided model—trained entirely on synthetic data—can generalize to real measurements and recover coherent demand and allocation structure despite differences in traffic dynamics, protocol behavior, and measurement noise. For each bottleneck capacity, we compare predicted per-user throughput against observed throughput over the first 200 windows, the first 400 windows, and the full trace, enabling a controlled evaluation of synthetic-to-real transfer under known capacities and
Table 5: Real-traffic evaluation of the TG-DIN on YouTube (𝑢 0 ) + Teams (𝑢 1 ) shared-link traces. 𝐶 (Mbps)
𝑛
20 60 120
835 722 759
RMSE𝑢0 RMSE𝑢1 RMSEmean MAE𝑢0 MAE𝑢1 2.16 3.58 4.47
1.05 0.52 0.25
1.60 2.05 2.36
1.52 1.40 1.82
0.89 0.41 0.16
competing real-world traffic. As shown in Table 5 and Fig. 5, TGDIN, albeit being trained entirely on synthetic data, generalizes effectively to real shared-link traces without any fine-tuning, recovering meaningful per-user allocation structure across all capacities. The transfer is especially strong for the dominant YouTube-like flow (𝑢 0 ), whose allocation level and temporal dynamics are consistently captured. At the tightest bottleneck of 20 Mbps, where contention is strongest, the model already achieves low absolute errors, with a mean RMSE of 1.60 Mbps. In this regime, minor deviations, such as slight over-estimation of the dominant flow and under-estimation of the smaller Teams-like flow (𝑢 1 ), are expected, given the pronounced short-term variability induced by severe bandwidth constraints. Importantly, these errors remain small relative to the link capacity. As capacity increases to 60 Mbps, alignment with the ground truth improves further. The model accurately captures the sustained throughput plateau of 𝑢 0 alongside the upward trend of 𝑢 1 , reducing the RMSE of the smaller flow to 0.52 Mbps while maintaining low absolute errors overall. At 120 Mbps, performance remains stable: although some bursty spikes in 𝑢 0 are slightly smoothed, the inferred trajectories closely follow the true throughput, and errors for 𝑢 1 become negligible (0.25 Mbps RMSE). Overall, mean RMSE grows only modestly with capacity, providing strong evidence that the theory-guided latent-demand representation learned from synthetic data transfers naturally to real-world traffic and enables stable, interpretable inference without target-side adaptation.
5
Conclusion
We presented TG-DIN, a theory-guided demand inference Network for latent demand inference from observable QoS signals. By introducing latent demand as an intermediate variable that explains observable network behavior through a differentiable theory layer, TG-DIN offers a more interpretable and mechanism-consistent alternative to direct black-box prediction. A key advantage of this
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
formulation is that the inferred latent demand is a practically meaningful representation of user-side network need, directly usable in downstream tasks such as congestion diagnosis, resource allocation, capacity planning, and policy evaluation. Experiments show that TG-DIN generalizes more robustly than purely data-driven baselines under cross-capacity shift, and preliminary real-traffic results suggest that the learned representation transfers beyond synthetic traces.
Acknowledgments This project is partially supported by the U.S. National Science Foundation under grant CNS-2344341.
References [1] Abd AlRhman AlQiam, Yuanjun Yao, Zhaodong Wang, Satyajeet Singh Ahuja, Ying Zhang, Sanjay G Rao, Bruno Ribeiro, and Mohit Tawarmalani. 2024. Transferable neural wan te for changing topologies. In Proceedings of the ACM SIGCOMM 2024 Conference. 86–102. [2] Giuseppina Andresini, Feargus Pendlebury, Fabio Pierazzi, Corrado Loglisci, Annalisa Appice, and Lorenzo Cavallaro. 2021. INSOMNIA: Towards ConceptDrift Robustness in Network Intrusion Detection. In Proceedings of the 14th ACM Workshop on Artificial Intelligence and Security (Virtual Event, Republic of Korea) (AISec ’21). Association for Computing Machinery, New York, NY, USA, 111–122. doi:10.1145/3474369.3486864 [3] Ons Aouedi, Van An Le, Kandaraj Piamrat, and Yusheng Ji. 2025. Deep learning on network traffic prediction: Recent advances, analysis, and future directions. ACM computing surveys 57, 6 (2025), 1–37. [4] Athula Balachandran, Vyas Sekar, Aditya Akella, Srinivasan Seshan, Ion Stoica, and Hui Zhang. 2013. Developing a predictive model of quality of experience for internet video. ACM SIGCOMM Computer Communication Review 43, 4 (2013), 339–350. [5] Giuseppe Bianchi. 2000. Performance analysis of the IEEE 802.11 distributed coordination function. IEEE Journal on Selected Areas in Communications 18, 3 (2000), 535–547. [6] Giovanna Carofiglio, Giulio Grassi, Enrico Loparco, Luca Muscariello, Michele Papalini, and Jacques Samain. 2021. Characterizing the relationship between application QoE and network QoS for real-time services. In Proceedings of the ACM SIGCOMM 2021 workshop on network-application integration. 20–25. [7] Dah-Ming Chiu and Raj Jain. 1989. Analysis of the increase and decrease algorithms for congestion avoidance in computer networks. Computer Networks and ISDN Systems 17, 1 (1989), 1–14. [8] Cisco Systems. 2012. Congestion Management Overview — IOS QoS Configuration Guide. Cisco Documentation. https://www.cisco.com/c/en/us/td/docs/ ios/qos/configuration/guide/12_2sr/qos_12_2sr_book/congstion_mgmt_oview. html [9] Fei Gui, Songtao Wang, Dan Li, Li Chen, Kaihui Gao, Congcong Min, and Yi Wang. 2024. RedTE: Mitigating subsecond traffic bursts with real-time and distributed traffic engineering. In Proceedings of the ACM SIGCOMM 2024 Conference. 71–85. [10] Chi-Yao Hong, Srikanth Kandula, Ratul Mahajan, Ming Zhang, Vijay Gill, Mohan Nanduri, and Roger Wattenhofer. 2013. Achieving high utilization with softwaredriven WAN. In Proceedings of the ACM SIGCOMM 2013 Conference on SIGCOMM. 15–26. [11] Shengxiang Hu, Guobing Zou, Bofeng Zhang, Shaogang Wu, Shiyi Lin, Yanglan Gan, and Yixin Chen. 2025. GACL: Graph Attention Collaborative Learning for Temporal QoS Prediction. IEEE Transactions on Network and Service Management (2025). [12] Rishabh Iyer, Katerina Argyraki, and George Candea. 2022. Performance interfaces for network functions. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22). 567–584. [13] Sushant Jain, Alok Kumar, Subhasree Mandal, Joon Ong, Leon Poutievski, Arjun Singh, Subbaiah Venkata, Jim Wanderer, Junlan Zhou, Min Zhu, et al. 2013. B4: Experience with a globally-deployed software defined WAN. ACM SIGCOMM Computer Communication Review 43, 4 (2013), 3–14. [14] Zeliang Kan, Feargus Pendlebury, Fabio Pierazzi, and Lorenzo Cavallaro. 2021. Investigating Labelless Drift Adaptation for Malware Detection. In Proceedings of the 14th ACM Workshop on Artificial Intelligence and Security (Virtual Event, Republic of Korea) (AISec ’21). Association for Computing Machinery, New York, NY, USA, 123–134. doi:10.1145/3474369.3486873 [15] Georgios Kougioumtzidis, Vladimir K Poulkov, Pavlos I Lazaridis, and Zaharias D Zaharis. 2025. Mobile network traffic prediction using temporal fusion transformer. IEEE Transactions on Artificial Intelligence (2025).
Yang et al.
[16] Jinsung Lee, Sungyong Lee, Jongyun Lee, Sandesh Dhawaskar Sathyanarayana, Hyoyoung Lim, Jihoon Lee, Xiaoqing Zhu, Sangeeta Ramakrishnan, Dirk Grunwald, Kyunghan Lee, et al. 2020. PERCEIVE: Deep learning-based cellular uplink prediction using real-time scheduling patterns. In Proceedings of the 18th International Conference on Mobile Systems, Applications, and Services. 377–390. [17] Shinan Liu, Francesco Bronzino, Paul Schmitt, Arjun Nitin Bhagoji, Nick Feamster, Hector Garcia Crespo, Timothy Coyle, and Brian Ward. 2023. Leaf: Navigating concept drift in cellular networks. Proceedings of the ACM on Networking 1, CoNEXT2 (2023), 1–24. [18] Harsha V Madhyastha, Tomas Isdal, Michael Piatek, Colin Dixon, Thomas Anderson, Arvind Krishnamurthy, and Arun Venkataramani. 2006. iPlane: An information plane for distributed services. In Proceedings of the 7th symposium on Operating systems design and implementation. 367–380. [19] Harsha V Madhyastha, Ethan Katz-Bassett, Thomas E Anderson, Arvind Krishnamurthy, and Arun Venkataramani. 2009. iPlane Nano: Path Prediction for Peer-to-Peer Applications.. In NSDI, Vol. 9. 137–152. [20] Alberto Medina, Nina Taft, Kave Salamatian, Supratik Bhattacharyya, and Christophe Diot. 2002. Traffic matrix estimation: Existing techniques and new directions. ACM SIGCOMM Computer Communication Review 32, 4 (2002), 161– 174. [21] K. Nichols, S. Blake, F. Baker, and D. Black. 1998. Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers. RFC 2474. IETF. https: //www.rfc-editor.org/rfc/rfc2474 [22] Abhay K. Parekh and Robert G. Gallager. 1993. A generalized processor sharing approach to flow control in integrated services networks: the single-node case. IEEE/ACM Transactions on Networking 1, 3 (1993), 344–357. [23] Yarin Perry, Felipe Vieira Frujeri, Chaim Hoch, Srikanth Kandula, Ishai Menache, Michael Schapira, and Aviv Tamar. 2023. { DOTE } : Rethinking (predictive) { WAN } traffic engineering. In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). 1557–1581. [24] Yan Qiao, Kui Wu, and Xinyu Yuan. 2024. AutoTomo: Learning-based traffic estimator incorporating network tomography. IEEE/ACM Transactions on Networking 32, 6 (2024), 4644–4659. [25] Mohammad Arafath Uddin Shariff, Venkat Sai Suman Lamba Karanam, and Byrav Ramamurthy. 2025. Traffic prediction for research and education networks using an ensemble GRU-LSTM with varying lead times. In 2025 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 1676–1681. [26] Paul Tune and Matthew Roughan. 2015. Spatiotemporal traffic matrix synthesis. In Proceedings of the 2015 ACM Conference on Special Interest Group on Data Communication. 579–592. [27] Bruno Missi Xavier, Magnos Martinello, Celio Trois, Brenno M. Alenca, and Ricardo A. Rios. 2024. Fast Learning Enabled by In-Network Drift Detection. In Proceedings of the 8th Asia-Pacific Workshop on Networking (Sydney, Australia) (APNet ’24). Association for Computing Machinery, New York, NY, USA, 129–134. doi:10.1145/3663408.3663427 [28] Shenghe Xu, Murali Kodialam, TV Lakshman, and Shivendra S Panwar. 2021. Learning based methods for traffic matrix estimation from link measurements. IEEE Open Journal of the Communications Society 2 (2021), 488–499. [29] Zhiying Xu, Francis Y Yan, Rachee Singh, Justin T Chiu, Alexander M Rush, and Minlan Yu. 2023. Teal: Learning-accelerated optimization of wan traffic engineering. In Proceedings of the ACM SIGCOMM 2023 Conference. 378–393. [30] Xinyu Yuan, Yan Qiao, Pei Zhao, Rongyao Hu, and Benchu Zhang. 2023. Traffic matrix estimation based on denoising diffusion probabilistic model. In 2023 IEEE Symposium on Computers and Communications (ISCC). IEEE, 316–322. [31] Bo Zhang, TS Eugene Ng, Animesh Nandi, Rudolf Riedi, Peter Druschel, and Guohui Wang. 2006. Measurement based analysis, modeling, and synthesis of the internet delay space. In Proceedings of the 6th ACM SIGCOMM conference on Internet measurement. 85–98. [32] Yazhuo Zhang, Rebecca Isaacs, Yao Yue, Juncheng Yang, Lei Zhang, and Ymir Vigfusson. 2023. Latenseer: Causal modeling of end-to-end latency distributions by harnessing distributed tracing. In Proceedings of the 2023 ACM Symposium on Cloud Computing. 502–519. [33] Yin Zhang, Matthew Roughan, Carsten Lund, and David Donoho. 2003. An information-theoretic approach to traffic matrix estimation. In Proceedings of the 2003 conference on Applications, technologies, architectures, and protocols for computer communications. 301–312. [34] Kevin Zhao, Prateesh Goyal, Mohammad Alizadeh, and Thomas E Anderson. 2023. Scalable tail latency estimation for data center networks. In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). 685–702. [35] Jie Zhou, Ding Ding, Ziteng Wu, and Yuting Xiu. 2023. Spatial context-aware time-series forecasting for QoS prediction. IEEE Transactions on Network and Service Management 20, 2 (2023), 918–931.
Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009