Adversarial Caching Training: Unsupervised Inductive Network Representation Learning on Large-Scale Graphs.

Area of Science:

• Data Mining
• Machine Learning
• Graph Theory

Background:

• Network representation learning (NRL) preserves graph structures in low-dimensional spaces for tasks like vertex classification.
• Graph convolutional network (GCN)-based models excel in inductive NRL, often using negative sampling (NS) for optimization.
• Current NS methods randomly select negative samples, leading to suboptimal or unrelated samples and potentially zero gradients, hindering representation learning.

Purpose of the Study:

• To propose an adversarial training method for unsupervised inductive NRL on large networks.
• To address the limitations of random negative sampling in GCN-based models.
• To improve the quality of learned network representations.

Main Methods:

• Developed an adversarial training method specifically for unsupervised inductive NRL.
• Designed a caching scheme with sampling and updating strategies for efficient, high-quality negative sample selection.
• Ensured the method is adaptive to various GCN-based models.

Main Results:

• The proposed method significantly improves performance in unsupervised inductive NRL.
• The caching scheme efficiently identifies high-quality negative samples, balancing exploration and training cost.
• Achieved better results compared to state-of-the-art models in extensive experiments.

Conclusions:

• The adversarial training method with an effective negative sampling strategy enhances unsupervised inductive NRL.
• The proposed approach offers a robust solution for learning superior graph embeddings on large-scale networks.
• This method provides a valuable advancement for data mining and machine learning applications involving graph data.