Fine-grained graph representation learning for heterogeneous mobile networks with attentive fusion and contrastive learning

TL;DR

This paper tackles learning dynamic, heterogeneous wireless data knowledge graphs (WDKGs) for network automation, where manual KG construction is costly and brittle. It introduces DMGSL, an unsupervised data-and-model driven graph structure learning framework that slices the network into homogeneous, time-sliced snapshots based on coherence time $T_c$ and fuses edge-type information with a hierarchical attention mechanism, while incorporating history via a temporal attention module built on LSTM. A contrastive learning component aligns anchor and learned graph representations under two augmented views, providing self-supervised guidance for structure refinement. Experiments on a real uplink throughput WDKG demonstrate that DMGSL outperforms state-of-the-art baselines in both learned topology quality and node classification performance, highlighting its potential for real-time network analytics and automation in 6G-era mobile networks.

Abstract

AI becomes increasingly vital for telecom industry, as the burgeoning complexity of upcoming mobile communication networks places immense pressure on network operators. While there is a growing consensus that intelligent network self-driving holds the key, it heavily relies on expert experience and knowledge extracted from network data. In an effort to facilitate convenient analytics and utilization of wireless big data, we introduce the concept of knowledge graphs into the field of mobile networks, giving rise to what we term as wireless data knowledge graphs (WDKGs). However, the heterogeneous and dynamic nature of communication networks renders manual WDKG construction both prohibitively costly and error-prone, presenting a fundamental challenge. In this context, we propose an unsupervised data-and-model driven graph structure learning (DMGSL) framework, aimed at automating WDKG refinement and updating. Tackling WDKG heterogeneity involves stratifying the network into homogeneous layers and refining it at a finer granularity. Furthermore, to capture WDKG dynamics effectively, we segment the network into static snapshots based on the coherence time and harness the power of recurrent neural networks to incorporate historical information. Extensive experiments conducted on the established WDKG demonstrate the superiority of the DMGSL over the baselines, particularly in terms of node classification accuracy.

Figures (7)

• Figure 1: Overall architecture of the proposed DMGSL. It consists of three modules: a) Hierarchical attention module (HAT). The input is the adjacency matrix of expert knowledge and feature matrix, the output is the anchor graph and learned graph integrating different types of edge; b) Temporal attention module (TAT). The input is the output of a) at different snapshots, the output is the anchor graph and learned graph integrating temporal information; c) Contrastive learning module. Calculate the contrastive loss after encoding and mapping the output of b), providing a self-supervised signal for unsupervised GSL.
• Figure 2: Schematic diagram of hierarchical attention module. The input is the adjacency matrix constructed by experts and the feature matrix of one snapshot. In the upper branch, the adjacency matrix is divided into three sub-adjacency matrices (causal/implicit/explicit relations) and mapped to a new feature space along with feature matrix. In the lower branch, the adjacency matrix is learned by a graph learner and then mapped to a new feature space along with feature matrix. The three anchor graphs and learned graphs are fused respectively through the edge-level attention mechanism.
• Figure 3: Heatmaps of adjacency matrices.
• Figure 4: Impact of mask rate on the classification performance.
• Figure 5: Performance with different model configurations.