Analysis of 5G academic Network based on graph representation learning method

Xiaoming Li, Guangquan Xu, Wei Yu, Pengfei Jiao, Xiangyu Song

Abstract

With the rapid development of 5th Generation Mobile Communication Technology (5G), the diverse forms of collaboration and extensive data in academic social networks constructed by 5G papers make the management and analysis of academic social networks increasingly challenging. Despite the particular success achieved by representation learning in analyzing academic and social networks, most present presentation learning models focus on maintaining the first-order and second-order similarity of nodes. They rarely possess similar structural characteristics of spatial independence in the network. This paper proposes a Low-order Network representation Learning Model (LNLM) based on Non-negative Matrix Factorization (NMF) to solve these problems. The model uses the random walk method to extract low-order features of nodes and map multiple components to a low-dimensional space, effectively maintaining the internal correlation between members. This paper verifies the performance of this model, conducts comparative experiments on four test datasets and four real network datasets through downstream tasks such as multi-label classification, clustering, and link prediction. Comparing eight mainstream network representation learning models shows that the proposed model can significantly improve the detection efficiency and learning methods and effectively extract local and low-order features of the network.

Introduction

The explosion of 5G technology has led to a massive increase in academic output, creating complex networks of collaborations, citations, and patents. Analyzing these academic social networks is vital for identifying key experts, predicting research trends, and fostering innovation. However, traditional graph representation learning models—which map high-dimensional network data into low-dimensional vectors—often struggle with the inherent sparsity of academic networks. Most existing methods focus on first-order (direct neighbors) and second-order (shared neighbors) similarities, failing to capture deeper structural patterns or "spatial independence" within the network.

Methodology

To address these limitations, we propose the Low-order Network representation Learning Model (LNLM). LNLM treats graph embedding as an optimization problem within an autoencoder-like framework, integrating two distinct feature extraction techniques:

  1. Local Structural Extraction: By applying Non-negative Matrix Factorization (NMF) to the network's adjacency matrix, the model captures immediate topological features.
  2. Low-order Feature Extraction: Utilizing Random Walks (inspired by DeepWalk), the model generates sequences that reveal deeper, "low-order" relationships between nodes that may not be directly connected.
  3. Joint Optimization: These features are fused into a single objective function. The model learns a representation matrix that balances local topology and broader random-walk patterns, controlled by hyperparameters.
Table 1: Mathematical symbols and definitions used in the LNLM framework.

The sliding window size ($T$) in the random walk process is a critical parameter. For dense networks, a small window suffices; however, for sparse academic networks, larger windows are necessary to uncover underlying community structures.

Datasets and Experimental Setup

The LNLM model was evaluated against eight mainstream baseline models across eight datasets:

  • Standard Benchmarks: Wikipedia, Polblog, Livejournal, and Orkut.
  • Collaboration Networks: GRQC, Hep-th, and Hep-ph from arXiv.
  • 5G-Specific Academic Networks: The Open Academic Graph (OAG) and a curated Academic Social Network (ASN) specifically filtered for 5G-related research, containing 2,407 nodes and 1,836 edges.

Performance was measured through multi-label classification, clustering (using Normalized Mutual Information, NMI), and link prediction (using Area Under Curve, AUC).

Key Results and Discussion

LNLM demonstrated superior performance across various tasks:

  • Clustering: On the Polblog dataset, LNLM achieved a 24% improvement in NMI compared to the next best model, highlighting its ability to detect community structures.
  • Link Prediction: On the Livejournal dataset, LNLM reached an AUC of 0.972, significantly outperforming traditional models like Node2vec (0.498) and LINE (0.515).
  • 5G Research Mapping: Applying LNLM to the 5G ASN identified 26 distinct research categories. "Network Intelligence" emerged as the largest category. The model also successfully identified central hubs in the 5G collaboration web, such as Mohsen Guizani.

The results suggest that the fusion of local topology and global sequence patterns is essential for analyzing sparse networks where information loss is a common issue for single-method models.

Conclusion and Future Work

The LNLM model effectively extracts local and low-order features, providing a robust tool for analyzing complex, sparse academic networks. While it shows significant promise, the authors note that academic networks are often multi-layered (e.g., citation vs. co-authorship layers). Future research will focus on extending LNLM to handle multi-layer sparse datasets more efficiently through localized optimization methods.

Summary updated

Abstract

With the rapid development of 5th Generation Mobile Communication Technology (5G), the diverse forms of collaboration and extensive data in academic social networks constructed by 5G papers make the management and analysis of academic social networks increasingly challenging. Despite the particular success achieved by representation learning in analyzing academic and social networks, most present presentation learning models focus on maintaining the first-order and second-order similarity of nodes. They rarely possess similar structural characteristics of spatial independence in the network. This paper proposes a Low-order Network representation Learning Model (LNLM) based on Non-negative Matrix Factorization (NMF) to solve these problems. The model uses the random walk method to extract low-order features of nodes and map multiple components to a low-dimensional space, effectively maintaining the internal correlation between members. This paper verifies the performance of this model, conducts comparative experiments on four test datasets and four real network datasets through downstream tasks such as multi-label classification, clustering, and link prediction. Comparing eight mainstream network representation learning models shows that the proposed model can significantly improve the detection efficiency and learning methods and effectively extract local and low-order features of the network.
BibTeX
@misc{li2021analysis,
  author = {Xiaoming Li and Guangquan Xu and Wei Yu and Pengfei Jiao and Xiangyu Song},
  title = {{Analysis of 5G academic Network based on graph representation learning method}},
  year = {2021},
  eprint = {2111.08264},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.08264},
}
Figure