Enhancing Graph Neural Networks with Adaptive Relation Learning via Variational Inference

Graph Neural Networks are highly effective in learning representations from graph-structured data. To better capture edge heterogeneity, prior approaches often partition edges using pre-defined rules or assign weights using attention mechanisms. However, these approaches suffer from critical limitations: pre-defined rules can be suboptimal, and deterministic methods fail to account for the inherent uncertainty in real-world graphs. To address these challenges, we propose the variational Relational Graph Convolutional Network (vRGCN), a novel framework that leverages Variational Inference to stochastically learn a distribution over latent relations directly from data. By employing a relation learner and a Gumbel-Softmax sampler, vRGCN adaptively infers the most effective relational structure for a given task without relying on any pre-defined information. Our experiments on graph and node classification benchmarks demonstrate that vRGCN consistently achieves superior performance against strong baselines. Furthermore, vRGCN is remarkably parameter-efficient, outperforming these baselines with significantly fewer parameters. Extensive ablation studies validate our design choices, confirming that stochastic sampling is crucial for robust exploration and that the ELBO objective is a vital regularizer for model generalization. vRGCN offers a flexible, powerful, and efficient paradigm for adaptive relation learning in graphs.