[Paper Review] Deep Adversarial Network Alignment

Tyler Derr, Hamid Reza Karimi+3 more|arXiv (Cornell University)|Feb 27, 2019

Advanced Graph Neural Networks19 references14 citations

TL;DR

This paper proposes Deep Adversarial Network Alignment (DANA), a novel unsupervised framework that aligns node embeddings across two networks using deep adversarial learning to match their underlying distribution, followed by efficient nearest-neighbor node alignment. DANA outperforms existing methods on real and synthetic datasets, achieving 54.2% and 51.1% accuracy on Bitcoin network variations, demonstrating superior performance without requiring seed nodes or side information.

ABSTRACT

Network alignment, in general, seeks to discover the hidden underlying correspondence between nodes across two (or more) networks when given their network structure. However, most existing network alignment methods have added assumptions of additional constraints to guide the alignment, such as having a set of seed node-node correspondences across the networks or the existence of side-information. Instead, we seek to develop a general network alignment algorithm that makes no additional assumptions. Recently, network embedding has proven effective in many network analysis tasks, but embeddings of different networks are not aligned. Thus, we present our Deep Adversarial Network Alignment (DANA) framework that first uses deep adversarial learning to discover complex mappings for aligning the embedding distributions of the two networks. Then, using our learned mapping functions, DANA performs an efficient nearest neighbor node alignment. We perform experiments on real world datasets to show the effectiveness of our framework for first aligning the graph embedding distributions and then discovering node alignments that outperform existing methods.

Motivation & Objective

• To develop a general, unsupervised network alignment method that does not rely on additional constraints such as seed nodes or side information.
• To address the challenge of aligning node embeddings from two separate networks, which are typically learned independently and thus not aligned in representation space.
• To improve node alignment performance by first aligning the global embedding distributions of two networks using adversarial training.
• To provide an unsupervised model selection heuristic that also validates the effectiveness of embedding distribution alignment.
• To demonstrate that deep adversarial learning can effectively align complex network structures even under noise and without supervision.

Proposed method

• DANA employs a two-stage framework: first, it uses a generative adversarial network (GAN) to learn a deep mapping from one network’s embedding space to the other’s, minimizing distributional discrepancy.
• The generator network learns to transform embeddings from one network to match the distribution of the target network’s embeddings, while the discriminator distinguishes real from generated embeddings.
• The adversarial training process ensures that the generated embeddings are indistinguishable from real embeddings of the target network, effectively aligning the two embedding distributions.
• After distribution alignment, DANA performs efficient nearest-neighbor node alignment in the shared embedding space using cosine similarity or other distance metrics.
• An unsupervised heuristic is introduced for model selection, based on the discriminator’s performance and the stability of the learned mapping.
• The framework is trained end-to-end using stochastic gradient descent, with loss functions designed to minimize the adversarial objective and maximize alignment quality.

Experimental results

Research questions

• RQ1 Can deep adversarial learning effectively align the embedding distributions of two independently learned networks without any additional constraints?
• RQ2 Does aligning the global embedding distributions lead to improved downstream node alignment performance compared to existing unsupervised methods?
• RQ3 How robust is the proposed DANA framework to structural noise, such as edge removal, in real-world networks?
• RQ4 Can DANA outperform state-of-the-art methods like REGAL-s2v, NetAlignBP, and IsoRank in terms of alignment accuracy without using side information or seed pairs?
• RQ5 Is there a reliable unsupervised method for selecting hyperparameters in the absence of ground truth alignment?

Key findings

• DANA achieved 54.2% alignment accuracy on the BitcoinA and BitcoinO dataset pair, significantly outperforming the next best baseline, REGAL-s2v, which achieved 12.4%.
• On the BitcoinAn and BitcoinOn pair, DANA achieved 51.1% accuracy, again surpassing all other baselines, including NetAlignBP (14.1%) and SparseIsoRank (4.7%).
• In experiments with pseudo-ground truth datasets (CollegeMsg, Hamsterster, Blogs), DANA maintained strong performance even under 20% edge removal, while all baselines failed almost completely at this noise level.
• The results indicate that DANA’s use of deep adversarial learning for embedding distribution alignment leads to more robust and accurate node alignment than shallow or spectral-based methods.
• The unsupervised model selection heuristic effectively identified optimal configurations and validated the importance of embedding distribution alignment as a key enabler of performance.
• DANA’s performance superiority demonstrates that advanced network embedding methods combined with adversarial training can significantly outperform traditional spectral or matrix factorization-based alignment techniques.