Can Classic Graph Autoencoders Rival SOTA? Surprising Optimizations Reveal Their Power

Link prediction remains a core problem in graph learning, underpinning applications such as recommendation systems and knowledge‑graph construction. While recent years have seen a surge of increasingly complex graph neural network (GNN) architectures, the authors observed that many reported gains stem from heavily tuned baselines rather than genuine architectural advances.

Motivated by this evaluation bias, the team revisited the classic Graph Autoencoder (GAE), a model introduced over a decade ago. They asked: if GAE were equipped with modern optimization tricks—linear convolutions, orthogonal initialization, careful hyper‑parameter tuning—how far could its performance be pushed without altering its simple core structure?

Through extensive ablation studies, the researchers identified a suite of techniques that consistently improve GAE across multiple datasets. These include:

Replacing the original inner‑product decoder with a more expressive linear convolutional decoder.

Applying orthogonal noise to node embeddings to preserve common‑neighbor information.

Systematic hyper‑parameter searches for learning rate, weight decay, and hidden dimension.

Optimized training pipelines that eliminate data leakage and ensure fair evaluation.

When evaluated on the large‑scale Open Graph Benchmark dataset ogbl-ppa, the optimized GAE achieved the #1 rank on the leaderboard, matching or surpassing the performance of many recent, far more complex models. Moreover, the streamlined architecture delivered speedups of tens to hundreds of times compared to those heavyweight baselines, making it practical for industrial‑scale graphs with billions of edges.

From an application perspective, these results suggest that efficiency and accuracy need not be mutually exclusive in link‑prediction tasks. The optimized GAE’s low computational overhead is especially valuable for real‑world recommendation systems, where massive item‑relation graphs demand fast inference.

The work was initially prepared as a research report, later refined into a paper submitted to the Conference on Information and Knowledge Management (CIKM). It received strong acceptance and positive reviewer feedback, highlighting the significance of revisiting and thoroughly optimizing foundational models.

Looking ahead, the authors plan two research directions: extending the optimized GAE to dynamic graph settings, where the graph structure evolves over time, and leveraging the insights gained to inform the design of more universal graph foundation models capable of handling diverse downstream tasks.