Goal-Oriented Learning at the Edge: Graph Neural Networks Over-the-Air for Blockage Prediction

Abstract

Sixth-generation (6G) wireless networks evolve from connecting devices to connecting intelligence. The focus turns to Goal-Oriented Communications, where the effectiveness of communication is assessed through task-level objectives over traditional throughput-centric metrics. As communication intertwines with learning at the edge, distributed inference over wireless networks faces a critical trade-off between task accuracy and efficient radio resource use. Traditional communication schemes (e.g., OFDMA) are not designed for this trade-off, often facing challenges related to scalability and latency. Therefore, we propose a novel goal-oriented framework that integrates over-the-air computation with spatio-temporal graph learning. Leveraging the wireless channel as an analog aggregation layer, the proposed framework enables low-latency message passing while efficiently aggregating semantically relevant features from distributed nodes. Theoretical analysis confirms that our analog architecture converges to the expressive power of digital message passing, while offering decisive scalability advantages. We assess the framework in proactive line-of-sight blockage prediction for millimeter-wave networks. Through high-fidelity ray-tracing simulations, the framework exhibits strong inductive generalization to unseen networks and adapts to domain shifts via lightweight transfer learning, matching or even outperforming digital baselines with significantly reduced communication overhead.

Figures (10)

• Figure 1: Example scenario illustrating the dependence of node cooperation on the prediction horizon $\tau$. (a) At $t=0$, Node 1 detects a mobile blocker while Nodes 2 and 3 maintain LOS. (b) For a short prediction horizon ($\tau=t+1$), Node 2 is informed by Node 1 to predict the incoming blockage. (c) For a longer horizon ($\tau=t+2$), the relevance of information shifts; Node 3 must now cooperate with Node 1 to anticipate the blocker's arrival.
• Figure 2: The network contains a set of wireless nodes $\mathcal{N}$ served by a central AP. The direct links are susceptible to blockage by dynamic obstacles, leading to NLOS conditions. To anticipate these events, nodes cooperate by exchanging local features $\mathbf{x}_{i,t}$ over $K$ parallel dynamic communication graphs $\{\mathcal{G}_{t,k}\}_{k=1}^K$, leveraging AirComp to aggregate information, and predicting the future blockage status $y_{i, t+\tau}$.
• Figure 3: Overview of the proposed GO-ST-AirGNN architecture. (Left) At the transmitter, node $i$ uses a semantic encoder $\Phi_i$ and a resource manager $\Psi_i$ to jointly generate the message embedding $\mathbf{m}_i$ and the goal-oriented power allocation vector $\mathbf{p}_i$. (Center) The wireless channel acts as a computation layer (AirComp), naturally aggregating neighbor signals via electromagnetic superposition across the parallel dynamic communication graphs. (Right) At the receiver, the semantic decoder $\chi_i$ processes the aggregated analog signals $\tilde{\mathbf{y}}_{i,k}$ together with the local features to predict the future blockage status $\hat{y}_{i, t+\tau}$.
• Figure 4: Digital twin of the BI-REX Competence Center simulated in NVIDIA Sionna RT. The network consists of distributed IIoT nodes (green spheres) served by a central AP (red sphere). A mobile robot acts as a dynamic blocker, traversing the factory floor and temporarily obstructing the direct LOS communication links between devices and AP.
• Figure 5: Performance evaluation of inductive generalization on unseen graph topologies.