MUST: Multi-Scale Structural-Temporal Link Prediction Model for UAV Ad Hoc Networks

Cunlai Pu, Fangrui Wu, Rajput Ramiz Sharafat, Guangzhao Dai, Xiangbo ShuarXiv 2025 · DOI · arXiv

Abstract

Predicting future communication links in UAV Ad Hoc Networks (UANETs) is a critical challenge due to high node mobility and extreme network sparsity. This paper introduces MUST (Multi-Scale Structural-Temporal), a deep learning framework that captures network features at micro (individual), meso (community), and macro (global) scales. By integrating these multi-scale features with a Stacked LSTM for temporal modeling and a custom sparsity-aware loss function, MUST significantly outperforms state-of-the-art methods. Experimental results across four mobility models demonstrate improvements in AUC and dramatic gains in AUPRC, highlighting the model's effectiveness in sparse aerial environments.

Predicting links in a network—knowing which nodes will connect in the next time step—is a classic challenge. However, doing this for UAV Ad Hoc Networks (UANETs) is particularly difficult. Unlike social networks or static grids, drone networks suffer from two major issues:

  1. High Dynamics: UAVs move at high speeds, causing links to establish and break constantly due to distance, obstacles, or weather.
  2. Extreme Sparsity: Drones are often spread over vast areas, meaning most potential links do not exist. Standard models often struggle to find meaningful patterns in such "empty" data.

Most existing models focus on a single structural scale (e.g., just looking at a node’s immediate neighbors) and fail to account for the unique "link quality" (signal-to-noise ratio) inherent in aerial wireless communication. Figure 1 illustrates the core task: predicting the future connectivity state $\hat{G}_l$ from a sequence of historical snapshots.

Temporal snapshots and prediction task in UANETs.

Methodology: The MUST Framework

The authors propose the MUST model, which breaks the problem down into three interconnected modules designed to handle the structural and temporal complexities of drone swarms.

Multi-Scale Feature Extraction

This is the core innovation of the framework. Instead of just looking at individual drones, the model extracts features at three distinct levels:

  • Micro-scale: Uses a Weighted Graph Attention Network (GAT) to learn features of individual UAVs based on their neighbors and link weights derived from distance and Signal-to-Noise Ratio (SNR).
  • Meso-scale: Uses the Louvain algorithm to detect drone "communities" (squads) and pools micro-features to see how these groups evolve collectively.
  • Macro-scale: Pools features from the entire network to capture global topology and overall network density.

Temporal Feature Extraction and Decoding

These multi-scale snapshots are fed into a Stacked LSTM (Long Short-Term Memory) network. This allows the model to "remember" how the movement of a community or an individual drone over previous snapshots influences their future position. Finally, a decoding module generates the predicted adjacency matrix $\hat{A}^l$.

The MUST model architecture, including multi-scale sub-modules and temporal feature extraction.

Sparsity-Aware Loss Function

To handle the fact that most links are zeros, the authors designed a custom loss function. It uses a weighted $L_2$ distance to prioritize correctly predicting existing links and an $L_1$ regularization term to ensure the predicted adjacency matrix remains sparse, preventing the model from making "safe" but useless 0.5 probability guesses.

Datasets and Evaluation

Because real-world military or industrial UAV topology data is rarely public, the authors simulated four distinct mobility models to generate benchmarks: Random Walk (RW), Manhattan Grid (MG), Reference Point Group (RPG), and Gauss-Markov (GM). Each dataset consists of 80 snapshots with 100 UAVs per simulation.

Key Results

The MUST model was tested against heavyweights like EvolveGCN, DySAT, and E-LSTM-D:

  • Superior Accuracy: MUST achieved an AUC (Area Under Curve) of up to 0.9367, outperforming the second-best models by 2.1% to 10.7%.
  • Precision in Sparsity: The most dramatic gains were in AUPRC (Area Under the Precision-Recall Curve), which is a better metric for sparse data. MUST improved AUPRC by 14.3% to 60.9% over competitors.
  • Ablation Success: The authors proved that removing any one scale (micro, meso, or macro) hurt performance, confirming that multi-scale analysis is essential for capturing swarm dynamics.

Key Insights

  • The Power of "Meso": Drones rarely move in total isolation; they often perform tasks in groups. By explicitly modeling the "Meso-scale" (communities), the model captures the collective intent of a swarm, which is a much more stable predictor than individual flight paths.
  • Link Quality as a Feature: Instead of treating a link as a binary 1 or 0, the model uses a weighting formula based on SNR and distance. This provides the GAT with "soft" information about which links are robust and which are about to fail.
  • Fighting the "0.5" Trap: In sparse networks, models often get "lazy" and predict a 0.5 probability for everything to minimize average error. The inclusion of $L_1$ regularization forces the model to make a choice, leading to much sharper and more accurate predictions.

Limitations and Future Work

While highly effective, the MUST model involves significant computational overhead due to the simultaneous execution of Louvain community detection, GATs, and LSTMs. Future research may focus on optimizing the architecture for real-time execution on constrained UAV hardware. Additionally, the model currently assumes perfect observation of historical topology, which may not hold in adversarial environments where nodes could be hidden or spoofed.

Summary updated

Abstract

Link prediction in unmanned aerial vehicle (UAV) ad hoc networks (UANETs) aims to predict the potential formation of future links between UAVs. In adversarial environments where the route information of UAVs is unavailable, predicting future links must rely solely on the observed historical topological information of UANETs. However, the highly dynamic and sparse nature of UANET topologies presents substantial challenges in effectively capturing meaningful structural and temporal patterns for accurate link prediction. Most existing link prediction methods focus on temporal dynamics at a single structural scale while neglecting the effects of sparsity, resulting in insufficient information capture and limited applicability to UANETs. In this paper, we propose a multi-scale structural-temporal link prediction model (MUST) for UANETs. Specifically, we first employ graph attention networks (GATs) to capture structural features at multiple levels, including the individual UAV level, the UAV community level, and the overall network level. Then, we use long short-term memory (LSTM) networks to learn the temporal dynamics of these multi-scale structural features. Additionally, we address the impact of sparsity by introducing a sophisticated loss function during model optimization. We validate the performance of MUST using several UANET datasets generated through simulations. Extensive experimental results demonstrate that MUST achieves state-of-the-art link prediction performance in highly dynamic and sparse UANETs.
BibTeX
@misc{pu2025must,
  author = {Cunlai Pu and Fangrui Wu and Rajput Ramiz Sharafat and Guangzhao Dai and Xiangbo Shu},
  title = {{MUST: Multi-Scale Structural-Temporal Link Prediction Model for UAV Ad Hoc Networks}},
  year = {2025},
  eprint = {2505.09331},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.09331},
}
Figure