A team led by Wei-Suo Ma from Peking University has developed a general-purpose optimization framework for Graph Autoencoders (GAEs), significantly enhancing their performance in link prediction tasks and demonstrating that simple, well-optimized models can rival or surpass state-of-the-art complex architectures. The results are particularly striking: after applying modern optimization techniques, the GAE model—once considered outdated—achieves top-tier performance, including ranking #1 on the large-scale ogbl-ppa benchmark from Stanford’s Open Graph Benchmark. The study’s contributions are twofold. First, it establishes a highly optimized GAE variant that achieves SOTA (state-of-the-art) results, proving the immense potential of revisiting foundational models with contemporary methods. Second, it identifies a set of broadly applicable techniques that improve GAE-based link prediction models, offering practical guidance for future model design. “We were thrilled to receive an acceptance decision and two strong accept reviews,” Ma said, reflecting on the paper’s submission process. Reviewers praised the core insight: that carefully optimized simple models can outperform complex ones in link prediction. The work was also commended for its originality and methodological rigor. From an application standpoint, the most significant impact lies in demonstrating that high efficiency and strong performance can coexist—a crucial advancement for scaling graph neural networks (GNNs). After optimization, the GAE model achieves performance improvements with efficiency gains of tens to hundreds of times over previous models. This is particularly valuable in real-world scenarios such as modern recommendation systems, where graphs can contain billions of edges. In such cases, the computational overhead of complex models often makes deployment impractical. The GAE’s lightweight structure offers a compelling alternative, enabling scalable deployment without sacrificing accuracy. The research was motivated by a critical observation in the field: many recent models claim performance gains over baselines that are, in fact, outdated and poorly optimized. This creates a misleading benchmark, encouraging researchers to focus on architectural novelty rather than uncovering the latent potential of simpler models. The team chose to revisit GAE—a foundational model introduced nearly a decade ago—and asked a fundamental question: What happens if we apply modern optimization techniques to the original GAE architecture without altering its core design? Rather than inventing a new model, the team focused on systematic optimization. Ma, then an undergraduate intern under Professor Mu-Han Zhang and senior lab members, led the project. Initial ablation studies revealed that removing advanced components from GAE had minimal impact on performance, suggesting that the model’s true capabilities had been severely underestimated. Further investigation confirmed this hypothesis. Simple enhancements—such as linear convolutions, proper parameter tuning, and careful data preprocessing—already delivered strong results, even outperforming some current baselines. The team rigorously verified that no data leakage or other artifacts were present, validating the robustness of their findings. To unlock the full potential, Ma and his team analyzed the source code of numerous leading link prediction models, extracting best practices and integrating them into the GAE framework. They designed large-scale experiments to identify optimal configurations for each component, culminating in a comprehensive, reproducible optimization pipeline. Theoretical analysis was also key. One major criticism of GAE has been its limited expressive power—its inability to capture structural signals critical for link prediction. The team discovered that when GAE uses orthogonal initialization for node embeddings, linear propagation, and dot-product prediction, it effectively preserves information about common neighbors—a vital structural signal. This explains the model’s strong performance and shows that its success is not accidental but rooted in sound design principles. The project initially began as a side investigation into negative sampling techniques for link prediction. The team’s earlier work highlighted efficiency gains from such methods, inspiring a broader pursuit of the balance between performance and computational cost. The NeurIPS paper that followed this research provided external validation, reinforcing the team’s confidence in the approach and encouraging them to submit their findings to the Conference on Information and Knowledge Management (CIKM). Throughout the writing process, the team refined their narrative, drawing structural inspiration from their NeurIPS paper. The final submission was well-received, with reviewers acknowledging the depth, clarity, and significance of the work. Ma emphasized the importance of deep foundational understanding and a patient, observational mindset. He warned that chasing complexity without thoroughly analyzing simpler models risks missing valuable design space and hinders true innovation. Looking ahead, the team plans to extend their work in two directions. First, they aim to adapt their optimization framework to dynamic graphs—where the network evolves over time—critical for real-world applications like recommendation systems. Second, they intend to explore the design of graph foundation models: unified, versatile architectures capable of handling diverse downstream tasks. The insights from this study are expected to provide valuable guidance for building such models. For more details, see the paper: https://arxiv.org/pdf/2411.03845 Edited and formatted by He Chenlong.