[Paper Review] Relation Learning on Social Networks with Multi-Modal Graph Edge Variational Autoencoders

Carl Yang, Jieyu Zhang+6 more|arXiv (Cornell University)|Nov 4, 2019

Advanced Graph Neural Networks39 references4 citations

TL;DR

This paper proposes ReLearn, a multi-modal graph edge variational autoencoder that jointly models latent relation distributions on social network links using graph convolutional networks and multiple reconstruction heads for text, attributes, and diffusion signals. It achieves superior performance in unsupervised relation learning, with robustness to noisy and incomplete data, and enables interpretable relation discovery through generative sampling and multi-decoder decoding.

ABSTRACT

While node semantics have been extensively explored in social networks, little research attention has been paid to profile edge semantics, i.e., social relations. Ideal edge semantics should not only show that two users are connected, but also why they know each other and what they share in common. However, relations in social networks are often hard to profile, due to noisy multi-modal signals and limited user-generated ground-truth labels. In this work, we aim to develop a unified and principled framework that can profile user relations as edge semantics in social networks by integrating multi-modal signals in the presence of noisy and incomplete data. Our framework is also flexible towards limited or missing supervision. Specifically, we assume a latent distribution of multiple relations underlying each user link, and learn them with multi-modal graph edge variational autoencoders. We encode the network data with a graph convolutional network, and decode arbitrary signals with multiple reconstruction networks. Extensive experiments and case studies on two public DBLP author networks and two internal LinkedIn member networks demonstrate the superior effectiveness and efficiency of our proposed model.

Motivation & Objective

• To address the challenge of learning implicit relation semantics on social network links, where relations are often under-specified and noisy.
• To integrate multi-modal signals—network proximity, user attributes, and information diffusion—into a unified framework for edge-level relation modeling.
• To develop a flexible, robust, and scalable model that operates effectively under limited or no supervision.
• To enable interpretable relation discovery by generating and decoding edge representations from learned latent distributions.

Proposed method

• The model uses a graph convolutional network (GCN) with neighborhood sampling to encode node and edge representations in large-scale networks.
• A Gaussian mixture variational autoencoder (GM-VAE) models the latent distribution of relation types per link, with global relation prototypes as Gaussian components.
• Two-step Monte Carlo sampling is employed for variational inference to estimate both global parameters and local relation distributions.
• Multiple dedicated decoders reconstruct signals from text, user attributes, and information diffusion, enabling multi-modal signal integration.
• The framework is trained end-to-end via reconstruction loss, allowing joint optimization of encoder and multiple decoders.
• Edge representations are generated via sampling from the learned mixture model, enabling interpretation and generation of relation-specific content.

Experimental results

Research questions

• RQ1 How can we effectively model multi-modal signals—network structure, user attributes, and information diffusion—for relation learning on social network links?
• RQ2 How can we learn a latent distribution of relations per link in the presence of noisy and incomplete data?
• RQ3 Can a unified generative model jointly reconstruct diverse modalities while maintaining robustness to attribute noise and missing links?
• RQ4 How can we interpret the learned latent relations through generative sampling and multi-decoder decoding?

Key findings

• ReLearn achieves state-of-the-art performance on relation learning tasks across two public DBLP networks and two internal LinkedIn networks, outperforming strong baselines in both link prediction and relation classification.
• The model demonstrates strong robustness to attribute noise, maintaining high performance even when up to 50% of attribute vectors are corrupted with unit Gaussian noise.
• ReLearn remains effective under link removal, showing stable performance when 2% to 10% of existing links are randomly removed from the network.
• Visualization of learned edge representations using PCA shows clear clustering of relation types (e.g., colleagues and schoolmates), confirming the model’s ability to disentangle distinct relations.
• Generative sampling from the learned Gaussian mixture model and decoding via multiple heads reveal interpretable content patterns: e.g., 'Management' and 'Performance' for colleagues, 'Learning' and 'Advising' for schoolmates.
• The runtime of ReLearn is comparable to efficient baselines like PTE and GraphSage, and significantly faster than heavier models like STNE and Inf2vec, indicating scalability.