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

TL;DR

This paper introduces a robust multidimensional graph neural network framework with edge-graph information bottleneck for improved signal processing in 6G wireless networks, enhancing spectrum efficiency and robustness against interference.

Contribution

It proposes a novel MDGNN architecture with hyper-edges and EGIB to effectively distinguish features and reduce irrelevant information, addressing limitations of traditional GNNs in wireless signal processing.

Findings

Achieves superior spectrum efficiency compared to existing methods.

Demonstrates increased robustness in interference-prone environments.

Maintains stable performance as interference noise increases.

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.