[Paper Review] MEGAN: A Generative Adversarial Network for Multi-View Network Embedding

Yiwei Sun, Suhang Wang+3 more|arXiv (Cornell University)|Aug 20, 2019

Advanced Graph Neural Networks17 references4 citations

TL;DR

MEGAN proposes a novel Generative Adversarial Network (GAN) framework for multi-view network embedding that jointly models connectivity within individual views and cross-view correlations to learn low-dimensional, non-linear, information-preserving node representations. Experiments show MEGAN outperforms state-of-the-art methods in node classification, link prediction, and visualization on real-world multi-view networks like Flickr and Last.fm.

ABSTRACT

Data from many real-world applications can be naturally represented by multi-view networks where the different views encode different types of relationships (e.g., friendship, shared interests in music, etc.) between real-world individuals or entities. There is an urgent need for methods to obtain low-dimensional, information preserving and typically nonlinear embeddings of such multi-view networks. However, most of the work on multi-view learning focuses on data that lack a network structure, and most of the work on network embeddings has focused primarily on single-view networks. Against this background, we consider the multi-view network representation learning problem, i.e., the problem of constructing low-dimensional information preserving embeddings of multi-view networks. Specifically, we investigate a novel Generative Adversarial Network (GAN) framework for Multi-View Network Embedding, namely MEGAN, aimed at preserving the information from the individual network views, while accounting for connectivity across (and hence complementarity of and correlations between) different views. The results of our experiments on two real-world multi-view data sets show that the embeddings obtained using MEGAN outperform the state-of-the-art methods on node classification, link prediction and visualization tasks.

Motivation & Objective

• Address the gap in multi-view network representation learning by integrating information from multiple network views with non-linear, low-dimensional embeddings.
• Overcome the challenge of modeling both intra-view connectivity and inter-view correlations in multi-view networks, which is absent in single-view GANs.
• Develop a GAN-based framework that generates adversarial samples to improve the robustness and representational quality of learned embeddings.
• Enable unsupervised training of multi-view network embeddings without requiring labeled data for downstream tasks.
• Demonstrate the superiority of the proposed method over existing state-of-the-art approaches across multiple evaluation tasks.

Proposed method

• Design a generator network that maps a shared latent space to adversarial node embeddings, learning to reconstruct multi-view network structures.
• Construct a discriminator that distinguishes between real multi-view network samples and adversarial samples generated by the generator, using both intra-view and inter-view connectivity patterns.
• Train the generator and discriminator in an adversarial manner using a minimax objective, where the generator aims to fool the discriminator by producing realistic multi-view network structures.
• Integrate multi-view information by jointly optimizing the generator to preserve connectivity in each individual view while capturing cross-view correlations.
• Use a contrastive loss to align the latent representations across views, enhancing the model’s ability to learn complementary features.
• Apply t-SNE for visualization of learned embeddings to assess clustering quality and structural coherence.

Experimental results

Research questions

• RQ1 Can a GAN-based framework effectively learn low-dimensional, non-linear embeddings for multi-view networks by modeling both intra-view connectivity and inter-view correlations?
• RQ2 How does the integration of complementary information across multiple views improve the quality of network representations compared to single-view or multi-view baselines?
• RQ3 To what extent does MEGAN outperform existing state-of-the-art methods in node classification, link prediction, and visualization tasks on real-world multi-view networks?
• RQ4 How sensitive is MEGAN’s performance to the choice of embedding dimension, and what optimal dimension yields the best results?
• RQ5 Does the adversarial training scheme in MEGAN enhance robustness and generalization in the presence of noisy or incomplete network structures?

Key findings

• MEGAN achieves state-of-the-art performance in node classification, with optimal performance at an embedding dimension of 128 on both Flickr and Last.fm datasets.
• On link prediction, MEGAN outperforms its single-view GAN counterpart (GraphGAN) and other multi-view methods, demonstrating the effectiveness of cross-view information integration.
• In visualization tasks using t-SNE, MEGAN-generated embeddings produce tighter, better-separated clusters for user interest categories compared to other methods.
• The performance of MEGAN is robust across different views, with consistent improvements over baselines even when individual views vary in link density and reliability.
• The model shows significant performance gains in scenarios where multiple views provide complementary information, especially when no single view contains all necessary information for accurate prediction.
• The ablation study confirms that jointly modeling intra-view and inter-view structures leads to better representation learning than modeling views in isolation.