Graph Neural Networks Powered by Encoder Embedding for Improved Node Learning

TL;DR

This paper leveraging a statistically grounded one-hot graph encoder embedding (GEE) as a high-quality, structure-aware initialization for node features yields the GEE-powered GNN (GG) framework, which provides consistent and substantial performance gains in both unsupervised and supervised settings.

Abstract

Graph neural networks (GNNs) have emerged as a powerful framework for a wide range of node-level graph learning tasks. However, their performance typically depends on random or minimally informed initial feature representations, where poor initialization can lead to slower convergence and increased training instability. In this paper, we address this limitation by leveraging a statistically grounded one-hot graph encoder embedding (GEE) as a high-quality, structure-aware initialization for node features. Integrating GEE into standard GNNs yields the GEE-powered GNN (GG) framework. Across extensive simulations and real-world benchmarks, GG provides consistent and substantial performance gains in both unsupervised and supervised settings. For node classification, we further introduce GG-C, which concatenates the outputs of GG and GEE and outperforms competing methods, achieving roughly 10-50% accuracy improvements across most datasets. These results demonstrate the importance of principled, structure-aware initialization for improving the efficiency, stability, and overall performance of graph neural network architecture, enabling models to better exploit graph topology from the outset.

Figures (7)

• Figure 1: A schematic diagram of the three algorithms for the node clustering task.
• Figure 2: Visualization of Training Dynamics and Clustering Quality. This figure contrasts the training process of our proposed GG (top row) and a vanilla GNN (bottom row) on a Degree-Corrected Stochastic Block Model (DC-SBM) graph (see Subsection \ref{['sec: Clustering Simulation']} for detailed settings). It demonstrates that GG's warm-start leads to a better initial state, faster convergence, and a superior final clustering quality. GG begins with a structured embedding (ARI = 0.05) and converges in only 1,800 epochs to a high-quality partition (ARI = 0.32). In contrast, the GNN starts from a random state (ARI $\approx$ 0.00) and requires a much longer training of 8,180 epochs to achieve an inferior result (ARI = 0.26).
• Figure 3: Clustering performance comparison of GEE, GNN and GG on DC-SBM graphs. This figure shows that our proposed GG offers the best trade-off between clustering accuracy (ARI, left) and speed (Time, right). While GEE is fastest, GG matches or surpasses its accuracy, especially on more challenging graphs ($r > 0.08$). Compared to the vanilla GNN, GG is both more accurate and substantially faster across all conditions.
• Figure 4: Classification performance comparison of GEE, GNN, GG and GG-C on DC-SBM graphs with 5% (top row) and 50% (bottom row) train/val set. This figure demonstrates the superiority of our GG and GG-C. They overcome the poor accuracy of the faster GNN (which is prone to local optima), while our concatenated variant, GG-C, also surpasses the classic GEE, particularly in the challenging low-label (5%) setting.
• Figure E1: Clustering performance comparison of GEE, GNN and GG on SBM graphs. This figure shows our GG method significantly outperforms the vanilla GNN in both accuracy and speed. Crucially, it also surpasses the classic GEE on more challenging graphs ($r > 0.08$), making it the most effective method overall.