WalkGAN: Network Representation Learning With Sequence-Based Generative Adversarial Networks.

Area of Science:

• Graph theory
• Machine learning
• Data science

Background:

• Network representation learning (or network embedding) aims to create low-dimensional vertex representations while preserving network structure.
• Real-world networks often have missing edges, which degrades the performance of existing methods that treat these as negative samples.
• Current approaches overlook the potential for true connections when edges are unobserved.

Purpose of the Study:

• To propose a novel network representation learning method, WalkGAN, that effectively captures true network structure, including unobserved links.
• To improve upon existing methods by inferring missing edges rather than treating them as non-existent.
• To enhance performance in downstream tasks like vertex classification, link prediction, and visualization.

Main Methods:

• WalkGAN integrates a random walk scheme with generative adversarial networks (GAN) into a network embedding framework.
• It utilizes GAN to generate synthetic vertex sequences that mimic network random walks.
• Vertex representations are learned from these generated sequences, inferring unobserved links with high probability.

Main Results:

• WalkGAN demonstrated significant performance improvements on benchmark network datasets.
• The method achieved better results in vertex classification tasks.
• Improvements were also observed in link prediction and network visualization tasks.

Conclusions:

• WalkGAN effectively infers unobserved links in networks by generating realistic random walk sequences.
• The proposed method offers a significant advancement in network representation learning.
• WalkGAN provides superior performance for vertex classification, link prediction, and visualization compared to existing techniques.