Gaussian Regularization in Neural Graph Learning

초록

Gaussian processes (GP), known for simplicity and flexibility, and the recent emergence of graph neural networks (GNNs) present promising avenues for semi-supervised learning on graph-structured data and beyond. Despite notable advancements in GNNs with a focus on neighborhood information, there are gaps in effectively integrating probability distributions and inter-entity relationships, as GNNs focus heavily on neighbors. Integrating probability distributions is essential to tackle noise and uncertainties. To address this issue, we present Gaussian Regularization in neural graph learning, which effectively incorporates Gaussian information to capture latent probabilistic distribution attributes from node embeddings based on the Gaussian Process, thereby boosting predictive capabilities. The process involves using a graph encoder to preprocess data, which is then encoded and used for effective aggregation and transformation of neighboring nodes. Gaussian Regularization works alongside any existing graph encoder model by encoding node representations into a Gaussian space to capture unique features. Logits are generated from this space, along with another set from an MLP, and then merged for final predictions. Extensive experiments using real-world benchmark datasets show that our approach outperforms several state-of-the-art GNN models and has a significant positive impact on off-the-shelf graph encoders and kernels, demonstrating its effectiveness and flexibility.

키워드