Title:Robust Multidimensional Graph Neural Networks for Signal Processing in Wireless Communications with Edge-Graph Information Bottleneck

Abstract:Signal processing is crucial for satisfying the high data rate requirements of future sixth-generation (6G) wireless networks. However, the rapid growth of wireless networks has brought about massive data traffic, which hinders the application of traditional optimization theory-based algorithms. Meanwhile, traditional graph neural networks (GNNs) focus on compressing inputs onto vertices to update representations, which often leads to their inability to effectively distinguish input features and severely weakens performance. In this context, designing efficient signal processing frameworks becomes imperative. Moreover, actual scenarios are susceptible to multipath interference and noise, resulting in specific differences between the received and actual information. To address these challenges, this paper incorporates multidimensional graph neural networks (MDGNNs) with edge-graph information bottleneck (EGIB) to design a robust framework for signal processing. Specifically, MDGNNs utilize hyper-edges instead of vertices to update representations to avoid indistinguishable features and reduce information loss, while EGIB encourages providing minimal sufficient information about outputs to avoid aggregation of irrelevant information. We numerically demonstrate that compared with existing frameworks, the proposed frameworks achieve excellent performance in terms of spectrum efficiency (SE) and network overhead under multiple signal processing tasks. Remarkably, as the interference noise increases, the SE performance of the proposed frameworks gradually stabilizes. This reveals the proposed frameworks have excellent robustness in interference prone environments, especially in wireless policies related to channel matrices.