Hub-Aware Sparsification of Interference Graphs for Learned Power Control

Graph neural networks (GNNs) are now a standard way to set transmit powers in dense wireless networks. Each transmitter–receiver link becomes a node, and each "who interferes with whom" channel becomes an edge. Message passing along those edges approximates the classical iterative optimizer in a few milliseconds. To scale to thousands of links, these models must drop most edges, and they do so with a single global rule: keep cross-links above a fixed interference level or within a fixed distance. Statistical graph learning has shown that one uniform cutoff fails on graphs with hubs. A cutoff strict enough to clean up the sparse periphery cuts off the hubs, and a cutoff loose enough to keep the hubs floods the periphery with noise. Real deployments have exactly this shape. A few high-power macro base stations interfere with almost everyone, while many low-power small cells interfere only locally.

We propose to prune the interference graph per receiver, against a budget on how much interference is thrown away. Transmitters that survive pruning at many receivers are separated out as "hub" nodes and handled by their own channel. Each receiver also gets one scalar feature: the interference it is no longer modeling edge by edge. The result is a bounded-degree peripheral graph plus a small hub set. We expect this to keep full-graph sum-rate at a small fraction of the edges and to keep the train-small, deploy-large generalization that wireless GNNs advertise, even in heterogeneous networks where uniform pruning breaks down.

Core Claim

Uniform interference pruning has the same structural flaw that statistical work identifies in uniform $\ell_1$ penalties. Two changes from that literature fix it, once they are restated in terms of what wireless performance depends on. The first is to calibrate the cutoff per column, as in Sparse Inverse Covariance Estimation with Calibration, which scales each column's penalty by its own norm. The second is to split hub structure from sparse structure, as in Learning graphical models with hubs, which decomposes the parameter matrix as $\Theta = Z + V + V^T$ with a sparse part $Z$ and column-structured hub part $V$.

The wireless restatement matters because performance does not depend on recovering individual edges. It depends on each receiver's signal-to-interference-plus-noise ratio (SINR), which is a sum over interferers. So the right pruning rule bounds the discarded interference mass at each receiver, not edge magnitudes.

For receiver $j$ with noise power $\sigma^2$, let $g_{ij}$ be the channel gain from transmitter $i$ and $p_{\max}$ the maximum power. We keep the smallest set $\mathcal{K}_j$ of strongest interferers such that

$$\sum_{i \notin \mathcal{K}_j,\, i \neq j} g_{ij}\, p_{\max} \;\le\; \epsilon \Big(\sigma^2 + \sum_{i \in \mathcal{K}_j} g_{ij}\, p_{\max}\Big).$$

At any power allocation, the SINR computed from kept edges then overestimates the true SINR by a factor of at most $1+\epsilon$. This follows directly because the discarded term is bounded by $\epsilon$ times the modeled denominator.

A transmitter that lands in more than $D$ of the kept sets is declared a hub. Its edges leave the sparse graph and go to a separate hub channel. This is the analogue of the group-sparse $V$ columns.

The residual feature is the actual discarded interference at the current powers. Physically, it is total measured received interference minus the modeled part, which a receiver can measure without per-link channel estimates. It is passed to each node so the model is told what pruning removed.

This construction also addresses a theoretical gap. Graph Neural Networks for Scalable Radio Resource Management proves size generalization for message-passing GNNs by tying them to distributed local algorithms, but only under a bounded number of interferers per node. Hub-dominated networks violate that assumption. Our peripheral graph has degree at most $D$ by construction, and the hub set is small, so the assumption is restored rather than assumed.

Relation to Prior Work

The statistical side motivates the proposal and supplies its mechanisms. The calibrated estimator above and reweighting schemes such as Learning Scale Free Networks by Reweighted $\ell_1$ regularization adapt penalties to node degree. Post-hoc thresholding in Thresholded Adaptive Validation removes leftover small entries and beats an oracle-tuned graphical lasso on hub graphs. All of these estimate a precision matrix from samples and score edge recovery. We borrow the per-column calibration and the hub/sparse split, but we score a downstream optimization through a physical SINR sum. That changes the right cutoff from "edge magnitude" to "aggregate mass."

The closest bridge is Learning Graph Structure from Convolutional Mixtures. It unrolls proximal gradient descent into a network with a single learned soft-threshold per layer, and its authors report difficulty on scale-free Barabási–Albert graphs. We read that failure as direct evidence of uniform sparsity bias inside learned, unrolled models. Our learned-threshold ablation tests whether per-node thresholds alone repair it.

On the wireless side, three works supply the models we sparsify. Unfolding WMMSE Using Graph Neural Networks (UWMMSE) unrolls the classical weighted minimum mean squared error (WMMSE) solver into four GNN layers. WCGCN is a max-pooling message-passing model. Random Edge GNNs (REGNN) apply polynomial filters of the raw channel matrix. All three were evaluated on homogeneous random geometric layouts. The REGNN authors state that it assumes node homogeneity and is sensitive to density shifts. We do not propose a new allocator. We propose a graph construction that these allocators run on, and we test them where their stated assumptions fail.

Cross-attention GNNs for skewed-degree GRNs show that hubs overwhelm ordinary attention in directed skewed graphs. That result constrains the obvious "just use attention" alternative, although it comes from gene regulatory networks rather than wireless.

Research Program

We will build a channel generator that extends the UWMMSE setup (Rayleigh fading, path-loss exponent 2.2, pairs dropped uniformly in 2D) with a heterogeneity knob. The knob controls two things: a number $M$ of macro transmitters at 20–30 dB higher power, and clustered small-cell placement. At $M=0$ the generator reproduces the published homogeneous regime exactly, which serves as a control.

On these channels we will first measure the trade-off that motivates the work. We sweep a uniform interference-to-noise cutoff, run WCGCN and UWMMSE on each pruned graph, and plot sum-rate against edges kept. The prediction is concrete. At $M=0$ the curve should have a clear knee. As $M$ grows, the knee should disappear, so that no single cutoff gets within 1% of full-graph sum-rate below roughly a third of the edges.

We then implement the budgeted, hub-split construction with $\epsilon$ and $D$ as its only two parameters, and pass the residual as a node feature. Both allocators are retrained on the pruned graphs. Finally, we test a learned variant in which $\epsilon$ is replaced by a per-receiver threshold produced by a small network. This shows whether learning adds anything once the budget rule is in place.

Baselines

The naive reference point has two parts: every transmitter at full power, and the full dense interference graph with no pruning. The first bounds how much power control is worth. The second bounds what pruning can at best preserve.

The strongest classical method is WMMSE run for 100 iterations. The best published learned results on the homogeneous regime are UWMMSE, with a reported sum-rate of 83.21 versus 82.94 for 100-iteration WMMSE on 20 pairs in the low-noise setting Unfolding WMMSE Using Graph Neural Networks, and WCGCN, reported at about 97–103% of WMMSE sum-rate with a 1–2% drop when trained on 50 pairs and deployed on 1,000 Graph Neural Networks for Scalable Radio Resource Management. Both allocators are run on full graphs, retrained on our heterogeneous generator, and given the same training steps, seeds, and hyperparameter search budget as our variant.

The pruning competitors are a uniform interference-to-noise cutoff, a distance cutoff, and per-receiver top-$k$. To compare at matched cost, each is tuned so that its average edges per receiver equal ours.

Two ablations isolate the key ingredients. One uses budgeted per-receiver pruning without hub separation. The other uses the full construction without the residual feature. REGNN is included as the homogeneity-assuming filter model.

Beating these baselines means three things. First, within 1% of full-graph sum-rate at no more than 10% of edges for $M>0$. Second, a gap over every matched-budget uniform rule that exceeds seed-to-seed variance. Third, a 50-to-1,000 generalization drop no larger than the full-graph model's under heterogeneity.

Why not just…?

Why not just keep the strongest $k$ interferers at each receiver? This is the most serious threat, because top-$k$ is already per-receiver and already gives bounded degree. Its weakness is that a fixed $k$ ignores how interference mass is spread. Near a macro station one interferer carries almost all the mass, while in a dense cluster dozens of comparable interferers do. A single $k$ therefore over-keeps at the first kind of receiver and under-keeps at the second. If top-$k$ combined with our residual feature matches the full construction at equal edge count, then the hub split and the budget rule are not needed. In that case we will report the residual feature itself as the contribution. The matched-budget ablation decides this.

Why not just prune by distance or a fixed interference level? These are the uniform rules the proposal targets. Map2Schedule reports that distance-based scheduling reaches only about 50% of achievable capacity in urban layouts. The threshold sweep in our first experiment directly tests whether a fixed-level rule has a usable operating point once hubs appear.

Why not just run attention over the full graph and let it learn which edges matter? Cost is quadratic in the number of links, which is about $10^6$ edges per layer at 1,000 links. That defeats the purpose of pruning. In addition, the evidence from Cross-attention GNNs for skewed-degree GRNs is that ordinary attention is itself dominated by hubs in skewed graphs. That evidence is from biology, so we include a full-graph attention model with a matched compute budget at $K\le 200$ to check it in the wireless setting.

Why not just run WMMSE? Published numbers already put it at about 16 ms for 20 pairs versus about 2 ms for UWMMSE, so it serves as the quality reference, not as the competitor.

Proposed Experiments

Training uses 50 pairs, and testing uses 6,400 channel draws at each of 50, 200, 500, and 1,000 pairs, with $M$ ranging from 0 to 8 macro transmitters. The metrics are sum-rate as a fraction of 100-iteration WMMSE, edges per receiver, GPU latency, the generalization drop from 50 to 1,000 pairs, and the empirical worst-case SINR overestimate, which should stay below $1+\epsilon$.

As a secondary, more realistic test, we use channel gains derived from RadioMapSeer urban radiomaps, where buildings create heterogeneity naturally. We compare against FPLinQ and Map2Schedule's reported "over 90%" of FPLinQ sum-rate on V2V scheduling, with the caveat that this is a scheduling task rather than continuous power control.

The decisive checks are three. First, at $M=0$ our method should give no gain over uniform pruning. If it does, the gain is not explained by hubs. Second, the advantage should grow with $M$. Third, the top-$k$-plus-residual ablation should fall measurably short. If any of these fails, the hub mechanism is not what drives the result.