Network operators face a basic tradeoff. The data that best explains what is going wrong, such as per-second radio indicators, per-flow traffic, and per-queue state, is also the most expensive data to collect, ship, and store across thousands of sites. What operators can afford to watch everywhere is coarse: averaged counters over minutes, link totals, a few summary indicators. This note proposes training the cheap, coarse monitor with help from an expensive, fine-grained model that exists only during training. The fine model teaches two things. It passes on its predictions, and it also shows which coarse measurements contain the moments and signals it found decisive. The containment relation is known exactly: each coarse counter is a known sum or average of fine quantities. That relation lets the fine model's "where to look" signal be mapped onto the coarse inputs.
The idea comes from computational pathology. There, models that read gigapixel slides at several magnifications improve when a fine-resolution branch teaches a coarse-resolution branch which tissue regions are critical. Each low-magnification patch physically contains its high-magnification children (A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification). Network telemetry has the same nesting, across time and across topology. Telecommunications differs from pathology in one way that changes the problem. The fine view is missing at deployment, not merely optional. So the question is how much of a fine-telemetry diagnostician's accuracy a coarse-telemetry monitor can recover, and whether the containment structure is what makes that recovery possible.
Core Claim
A coarse-input model will recover substantially more of a fine-input teacher's accuracy if two conditions hold. First, it is trained to match the teacher's predictions. Second, its attention over coarse inputs is aligned with the teacher's attention pooled through the exact fine-to-coarse containment map. We claim the gain will be concentrated on subtle, transitional faults, where averaging hides the signal. We also claim it will vanish if the containment map is scrambled. The second prediction is what separates this proposal from generic knowledge distillation.
The training objective for the coarse student $S$, given a fine teacher $T$, is
$$ \mathcal{L} = \mathcal{L}_{\text{task}}(S) + \alpha\, \mathrm{KL}\big(p_T \,\|\, p_S\big) + \beta \sum_{c} \Big( a_S(c) - \sum_{f \in c} a_T(f) \Big)^2 . $$
Here $p_T$ and $p_S$ are the two models' output distributions. $a_T(f)$ is the teacher's normalized importance on fine unit $f$, which could be a 3-second sample of one indicator or one flow's queue state. $a_S(c)$ is the student's normalized importance on coarse unit $c$, which could be a 30-second average or a link counter. The sum over $f \in c$ runs over the fine units that coarse unit $c$ physically aggregates. $\alpha$ and $\beta$ weight the two transfer terms. The third term is the new ingredient. It plays the role of the cross-scale "criticality" transfer in DAS-MIL, rebuilt for a setting where the child level is never observed at test time.
Where This Sits in Prior Work
Work that motivates the gap. The diagnosis system RAID: Root Cause Anomaly Identification and Diagnosis learns contrastive representations of normal Wi-Fi telemetry and attaches lightweight root-cause classifiers. It reaches 81.83% accuracy and 81.60% F1 on 112 indicators sampled every 3 seconds. It struggles with subtle, transitional degradations, especially coverage faults. Its fine sampling is exactly what a wide-area deployment cannot afford everywhere. We take RAID as the teacher architecture and ask what happens when its inputs are coarsened.
The network model RouteNet-Fermi: Network Modeling With Graph Neural Networks predicts path delay with 6.24% mean relative error by passing messages among flows, queues, and links. It assumes per-flow traffic descriptions as input. Operators often lack those and hold only link-level aggregate counters. The survey Network Digital Twin: Context, Enabling Technologies and Opportunities names flow-level fidelity at scale as an open problem. Our coarse student is one answer to that problem: it keeps flow-level knowledge in its weights rather than in its inputs.
Work that supplies the method. We reuse the self-distillation between resolution tiers from DAS-MIL (A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification). In DAS-MIL both resolutions are available at inference and distillation acts as a regularizer. In our setting the fine tier is privileged information seen only in training, so the containment map is the only bridge. We borrow the explicit parent–child containment idea from the tree model in Diagnose Like A Pathologist: Weakly-Supervised Pathologist-Tree Network for Slide-Level Immunohistochemical Scoring, which enforces parent–child consistency. We borrow the vertical cross-scale edges from A Multi-scale Graph Network with Multi-head Attention for Histopathology Image Diagnosisn. Those models propagate information along containment edges. We use containment only to route a training signal.
Work that constrains the design. Deep Learning for Visual Recognition of Environmental Enteropathy and Celiac Disease found that concatenating multi-zoom features often fails to beat single-scale models. This warns that stacking multiple resolutions is no guarantee of benefit, which is one reason we test a structured transfer rather than more inputs. DAS-MIL also reports that feature-similarity edges beat physical adjacency. That cautions against assuming physical containment is the right structure, and we test it directly with a learned-grouping control.
Adjacent work we do not build on. The cellular forecaster Adaptive Hybrid Spatial-Temporal Graph Neural Network for Cellular Traffic Prediction and the language-model planner Hierarchical Debate-Based Large Language Model (LLM) for Complex Task Planning of 6G Network Management are related but solve different problems. The first works on a single flat spatial scale. The second keeps its hierarchy in text. Neither addresses missing fine telemetry.
Research Program
The program has two testbeds, chosen so that coarsening happens along different axes: time and indicators in one, topology in the other.
Diagnosis on the RAID Wi-Fi testbed. The teacher is RAID trained on the full input: 10 steps of 3-second samples over 112 indicators per 30-second window, across the Normal, Coverage, and Interference scenarios. The student sees a coarsened version of the same window, and we vary the coarsening level. At the mildest level, each window is reduced to per-indicator mean, max, and min. At a stronger level, indicators are pooled into roughly a dozen group summaries such as signal strength, transmission performance, and channel utilization, mimicking the counters an operator would export. At the strongest level, several windows are merged into multi-minute bins. At every level the containment map is exact by construction. If the RAID data cannot be obtained, we will rebuild an equivalent labelled Wi-Fi fault set with the same scenario design. We will report this as a deviation.
Performance modeling on RouteNet data. We use the public RouteNet OMNeT++ setting: training on 25–50-node topologies and testing on 50–300-node topologies. The teacher is RouteNet-Fermi with its normal flow, queue, and link inputs. The student receives only per-link average utilization plus routing, with no per-flow traffic descriptors. The containment map sends each link counter to the set of flows routed over it. The teacher's importance is taken from its flow-level attention or gradient attribution. The student's importance is its link-level attention.
Shared analysis. In both settings we will measure the "recovery fraction". This is the share of the gap between a coarse model trained alone and the fine teacher that each method closes. We plot it against telemetry cost, measured in bytes per site per hour. This curve is the practical output: it tells an operator how much fine telemetry can be dropped for a given loss of diagnostic quality.
Baselines
The naive reference point is the student architecture trained only on coarse inputs and labels, with the same tuning budget as every other variant. It measures how much the coarse data supports on its own. Every distillation result is reported as a recovery fraction relative to this model and the teacher.
The ceiling is the fine teacher itself. On RAID, the reported figure is 81.83% accuracy and 81.60% F1 (RAID: Root Cause Anomaly Identification and Diagnosis). On RouteNet data, it is 6.24% delay mean relative error (RouteNet-Fermi: Network Modeling With Graph Neural Networks). We retrain both teachers ourselves and report our reproduced numbers next to these.
The strongest existing competitor is standard prediction-level distillation, which is our objective with $\beta = 0$. This is the obvious thing a practitioner would try, and it may already capture most of the benefit. For the RouteNet setting we add a second reference: queueing theory computed from link loads, which the RouteNet-Fermi paper reports reaches up to 70% mean relative error. This is the classical tool that needs only coarse inputs. On RAID we also train the paper's representation-learning baselines (TS2Vec, TS-TCC, and the others RAID lists) on coarse inputs. No reported result on coarsened RAID inputs was found on Lacuna, so these numbers will be new.
Fairness. All students share the architecture, data splits, number of training epochs, and a fixed hyperparameter search of equal size. The $\alpha$ and $\beta$ search counts against our budget.
Isolating the key ingredient. Two ablations test whether the containment term is doing the work. In the first, we keep the importance-alignment term but randomly permute which fine units map to which coarse unit, preserving group sizes. In the second, we replace physical containment with groups learned by clustering the teacher's fine-unit embeddings. This control speaks directly to DAS-MIL's finding that feature similarity beats physical adjacency.
What counts as beating the competitor. Beating prediction-level distillation means a higher recovery fraction on the Coverage class on RAID, and lower mean relative error on the larger, unseen RouteNet topologies. In both cases the gain must exceed seed-to-seed variation over at least five seeds, and it must disappear under the permuted map.
Why not just…?
Why not just train the coarse model directly on labels? That is the naive baseline above, and it may be enough for faults that show up as large shifts in averages, such as heavy interference. Our prediction is narrower: averaging erases short, transitional events. That is where RAID already struggles at full resolution, so coarse-only training should fail most on the Coverage class. If the coarse-only model matches the teacher across all classes, the proposal is unnecessary and the cost curve will show it.
Why not just use ordinary distillation on the teacher's outputs? This is the real threat. Prediction-level distillation transfers soft labels that may already encode most of what matters. The permuted-map ablation settles the question. If importance alignment helps only when the containment map is correct, then the hierarchy carries information beyond the soft labels. If permuted and correct maps help equally, the extra term is just a regularizer, and we would report that the simpler method suffices.
Why not just collect fine telemetry on demand when a coarse alarm fires? Operationally this is a strong alternative. We run it as a two-stage baseline: a coarse detector triggers fine collection for the flagged windows, and the teacher then diagnoses them. We place it on the same accuracy-versus-cost curve. It has two built-in weaknesses. It misses faults the coarse detector never flags, and it adds delay. Distillation could also improve the trigger itself, so the two may combine. We report the combined variant as well rather than presenting distillation as a replacement.
Why not just concatenate summaries at several time scales, or use a bigger flat spatio-temporal graph model? Feeding more coarse views into one model adds no information that the fine data held. The pathology evidence also suggests flat multi-zoom concatenation often fails to beat a single scale (Deep Learning for Visual Recognition of Environmental Enteropathy and Celiac Disease). We include a multi-window concatenation student as an extra baseline. It costs little and directly tests whether extra coarse views make up for missing fine ones.
Decisive Checks
Three results would make or break the proposal. The first is the per-class recovery fraction on RAID. The claim requires the importance-alignment term to raise Coverage-class F1 beyond prediction-only distillation, with smaller or no gains on Interference. The second is the RouteNet out-of-distribution test on 50–300-node topologies. The student trained only on link counters must close a meaningful part of the distance between queueing theory and RouteNet-Fermi's 6.24% error, and the alignment term must widen that margin on the largest topologies, where flow detail matters most. The third is the permuted-map control in both settings. If the gains survive scrambling, we drop the containment claim and report the outcome as a negative result about transferring multi-scale ideas from pathology to network monitoring.
References
- A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification
- RAID: Root Cause Anomaly Identification and Diagnosis
- RouteNet-Fermi: Network Modeling With Graph Neural Networks
- Network Digital Twin: Context, Enabling Technologies and Opportunities
- Diagnose Like A Pathologist: Weakly-Supervised Pathologist-Tree Network for Slide-Level Immunohistochemical Scoring
- A Multi-scale Graph Network with Multi-head Attention for Histopathology Image Diagnosisn
- Deep Learning for Visual Recognition of Environmental Enteropathy and Celiac Disease
- Adaptive Hybrid Spatial-Temporal Graph Neural Network for Cellular Traffic Prediction
- Hierarchical Debate-Based Large Language Model (LLM) for Complex Task Planning of 6G Network Management