🔢 mathematics

GNN-RSMA: An Interference Management Framework for a Large-Scale HAPS Network

This paper proposes GNN-RSMA, a scalable interference management framework for large-scale HAPS networks that leverages graph neural networks to optimize UE clustering and rate-splitting multiple access power allocation, thereby maximizing minimum spectral efficiency with significantly lower computational cost than traditional optimization methods.

Original authors: Afsoon Alidadi Shamsabadi, Animesh Yadav, Halim Yanikomeroglu

Published 2026-08-04

📖 4 min read🧠 Deep dive

CC0 1.0

Original authors: Afsoon Alidadi Shamsabadi, Animesh Yadav, Halim Yanikomeroglu

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the sky above our cities is about to get a whole lot busier. For decades, we've relied on cell towers on the ground to keep our phones connected, but as we add more devices—from smartphones to drones—the ground is getting crowded. Enter High Altitude Platform Stations (HAPS): think of them as giant, stationary platforms hovering high in the stratosphere, about 20 kilometers up. They act like floating cell towers, offering a bird's-eye view that covers a massive area, much wider than any tower on the ground.

The big challenge with these floating towers is interference. Because they are so high up, they can see almost everything below them with a clear, direct line of sight. This is great for signal strength, but it's a nightmare for organization. Imagine a teacher in a classroom trying to talk to 60 students at once. If everyone is shouting, no one can hear. In the HAPS world, if the tower tries to send data to many users at once using the same radio frequencies, the signals crash into each other, creating a chaotic mess of noise. To fix this, engineers use a clever trick called "clustering," where they group users together and assign them specific beams, like a spotlight focusing on a small group. They also use a technique called RSMA (Rate-Splitting Multiple Access), which is like a smart waiter who splits a complex order into a "common" part everyone hears and a "private" part just for you, managing the noise so everyone gets their message. But figuring out exactly how much power to give each signal to keep everyone happy is a math problem so hard that even the best computers take forever to solve it.

This paper, titled "GNN-RSMA," tackles that impossible math problem with a new kind of artificial intelligence. The authors, working with a network of high-altitude platforms, realized that the old way of solving the power-allocation puzzle was too slow for real-time use. They proposed a solution using a Graph Neural Network (GNN), which is a type of AI designed to understand relationships, like how friends in a social network are connected. Instead of trying to solve the math equation from scratch every time the wind blows or a user moves, the AI learns the "shape" of the problem. It treats users, signal beams, and radio frequencies as nodes in a giant, shifting graph.

The researchers simulated a network with 60 users and 10 radio channels to test their idea. They found that their AI system could make decisions about how to split the power and manage the interference almost instantly. In fact, it was about 40 to 60 times faster than the traditional, heavy-duty math method (called SCA) that is currently considered the gold standard. While the traditional method takes nearly 28 seconds to solve a single scenario with 100 users, the AI does it in less than half a second.

Crucially, the paper shows that this speed doesn't come at the cost of quality. The AI's solution was nearly as fair and efficient as the slow, perfect math method. It managed to keep the "worst-off" user (the one with the poorest signal) happy, ensuring that no one was left in the dark. The authors also discovered that the AI was smart enough to handle different numbers of users without needing to be retrained; if you added more people to the network, the same AI model just adapted on the fly.

However, the paper is careful to note that these results come from computer simulations, not a real-world test in the sky. The authors argue that while the AI is a massive step forward for speed and scalability, it is a proposed framework that needs real-world validation. They also explicitly highlight that simply adding more antennas to the HAPS isn't always the answer; their simulations showed that while larger antennas improve interference suppression up to a point, making them too large creates beams that are too narrow, which actually degrades service for users on the edge of the coverage area.

In the end, this paper suggests that by teaching an AI to "see" the network as a connected web rather than a list of numbers, we can manage the complex interference of future sky-high networks. It offers a way to keep our future connections fast, fair, and ready for the moment we launch thousands of these floating towers into the sky.

Technical Summary: GNN-RSMA for Large-Scale HAPS Networks

1. Problem Statement

The paper addresses the challenge of interference management in large-scale High Altitude Platform Station (HAPS) networks integrated with terrestrial infrastructure to form vertical heterogeneous networks (vHetNets). While HAPS offers wide coverage and low latency, its high-altitude deployment creates specific propagation characteristics:

Strong Line-of-Sight (LoS) Links: This leads to highly correlated channels among User Equipments (UEs), making it difficult to distinguish users via spatial beamforming alone.
Resource Scarcity: A single HAPS serves a large number of UEs with limited radio resources (Resource Blocks, RBs). Consequently, multiple UEs must share the same RB, causing severe intra-RB interference.
Computational Complexity: Traditional optimization methods, such as Successive Convex Approximation (SCA), can solve the Max-Min Fairness (MMF) power allocation problem but suffer from prohibitive computational complexity as the network scales. They require solving the optimization from scratch for every new network realization, making them unsuitable for real-time, large-scale deployment.

The core problem is to design a scalable, low-complexity interference management scheme that maximizes the minimum spectral efficiency (SE) among all UEs in a multi-cluster HAPS network using Rate-Splitting Multiple Access (RSMA).

2. Methodology

The proposed solution, GNN-RSMA, integrates UE clustering with RSMA and leverages Graph Neural Networks (GNNs) for power allocation.

A. System Model and Clustering

Clustering: The network employs a "Worst-UE" (WU) clustering algorithm. UEs are grouped into clusters based on angular proximity. The beam for each cluster is steered toward the UE with the worst channel condition within that cluster to maximize the minimum beam gain.
RSMA Framework: Within each cluster, UEs share orthogonal RBs. However, different clusters share the same RBs, creating inter-cluster interference. RSMA is used to manage this by splitting each UE's message into a common stream (decoded by all UEs on the RB) and a private stream (decoded only by the intended UE).
Optimization Goal: Jointly optimize the transmit powers for common and private streams ( $p^{(c)}_{r,\ell}$ and $p^{(p)}_u$ ) and the common rate allocation ( $C_u$ ) to maximize the minimum achievable rate ( $\min_u R_u$ ) subject to total power and decodability constraints.

B. Graph Representation

The HAPS network is modeled as a heterogeneous graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ to capture the complex interactions:

Node Types: Three distinct node types are defined: UEs, Clusters, and RBs.
Node Features:
- UEs: Include position, serving beam gain, max/mean interference gains.
- Clusters: Include centroid position, serving gain statistics, and leakage characteristics.
- RBs: Include statistics of UEs sharing the RB (min/max/mean gains) and cross-cluster interference potential.
Edge Types: Four relation types capture physical interactions: UE-to-UE (interference), UE-to-Cluster (serving/interfering), UE-to-RB (assignment), and Cluster-to-RB (usage).

C. GNN Architecture

The proposed GNN policy maps the graph representation directly to RSMA power allocation decisions without iterative optimization.

Heterogeneous Message Passing: The network uses Edge-Conditioned Graph Attention Networks (GAT) with multi-head attention. Different edge types (relations) have independent attention mechanisms to weigh neighbors based on physical link conditions (e.g., channel gain).
Global Context: To address the global nature of the MMF objective (which depends on the worst-performing UE), a context vector is extracted via pooling (mean, max, min) of UE embeddings and injected back into the UE representations.
Constraint-Feasible Decoding: Instead of using penalty terms in the loss function, a differentiable decoder enforces constraints by construction:
- A Softmax layer scales the output logits to ensure the sum of all common and private powers equals the total transmit power budget ( $P_T$ ).
- A per-RB Softmax distributes the common rate among UEs sharing that RB, ensuring decodability constraints are met.
Iterative Refinement: To emulate the iterative nature of SCA, the GNN is unrolled over $N_p$ refinement passes. The rate achieved in one pass is fed as a "rate hint" into the next pass, allowing the policy to progressively refine the allocation and identify bottleneck UEs.
Unsupervised Training: The model is trained without labeled data (no SCA solutions required). It maximizes a smooth surrogate of the MMF objective (log-sum-exp lower bound) over mini-batches of randomly generated network realizations.

3. Key Contributions

Heterogeneous GNN Policy: The paper proposes the first GNN-based framework tailored for MMF-RSMA power allocation in multi-cluster HAPS networks. It represents the network as a heterogeneous graph (UE, Cluster, RB nodes), allowing the policy to capture the underlying communication topology and generalize across different network sizes.
Constraint-Feasible Architecture: The design incorporates a decoder that maps network outputs to power allocations satisfying total power and common-rate decodability constraints by construction, eliminating the need for constraint penalties during training.
Scalability and Generalization: Due to the permutation-invariant nature of GNNs, a single trained policy adapts to networks with varying numbers of UEs, clusters, and RBs without retraining.
Iterative Refinement Mechanism: An unrolled inference process emulates the inner iterations of SCA algorithms, improving the identification of bottleneck UEs and refining power allocation with minimal computational overhead.

4. Simulation Results

The proposed GNN-RSMA was evaluated against an SCA-based RSMA benchmark, NOMA, and SDMA schemes in a HAPS network with 60 UEs and 10 RBs.

Performance: GNN-RSMA achieves approximately 92% of the Max-Min Fairness performance of the SCA benchmark (0.493 b/s/Hz vs. 0.536 b/s/Hz median SE). It significantly outperforms NOMA and SDMA.
Fairness: The algorithm achieves a Jain's fairness index median of 0.995, indicating highly equitable resource distribution among UEs.
Computational Efficiency: The inference time for GNN-RSMA is orders of magnitude faster than SCA. For 60 UEs, GNN-RSMA takes ~0.3 seconds, while SCA takes approximately 10.8–27.9 seconds (the range observed for 40 to 100 UEs). This gap widens as the number of UEs increases (e.g., >60x speedup at 100 UEs).
Generalization: A single model trained on variable-sized networks maintains stable performance across different UE counts (40 to 100), whereas the SCA performance degrades slightly as interference increases.
Clustering Impact: The WU-Clustering algorithm is shown to be critical; replacing it with centroid-based clustering causes a 63% drop in SEmin as UE count increases, whereas WU-Clustering maintains near-constant performance.

5. Significance and Claims

The paper claims that GNN-RSMA provides a scalable, practical alternative to convex-optimization-based interference management for real-time RSMA deployment in HAPS networks.

Real-Time Viability: By replacing iterative optimization with a few low-complexity forward passes, the method enables real-time resource allocation in dynamic, large-scale networks where SCA is too slow.
Robustness: The approach demonstrates that AI-driven methods can learn the complex mapping from network states to optimal resource allocations, generalizing effectively to unseen network sizes and topologies without retraining.
Interference Management: The study validates that combining RSMA with GNN-based power allocation effectively manages the severe intra-RB interference inherent in high-altitude, wide-coverage HAPS deployments, achieving fairness comparable to the theoretical optimization benchmark.