Introduction
Graph Contrastive Learning (GCL) has emerged as a powerful paradigm for learning node representations without manual labels. Most GCL methods rely on the InfoNCE loss, which works by pulling "positive" pairs (different views of the same node) together and pushing "negative" pairs (different nodes) apart. However, researchers have identified a fundamental characteristic of GCL that is often overlooked: it is essentially a $(2n-1)$-way 1-shot classification problem. In this setup, every node in the graph acts as its own unique "ideational class" (CI-class), and the model only ever encounters a single sample for that class.
The Limitations of Deterministic Embeddings
Standard GCL methods represent these CI-classes using deterministic vectors—single points in a high-dimensional space. With only one sample available per class, a single point is often insufficient to capture the potential distribution of that class. Furthermore, observed graph structures are frequently noisy or incomplete. Treating edges as binary "hard" connections ignores the underlying probability of these relationships, leading to a "limited self-supervised signal" that can hinder representation quality.
The OS-GCL Framework
To address these challenges, the OS-GCL (One-Shot Learner in Graph Contrastive Learning) framework replaces deterministic points with probability distributions for both node features and graph topology.
1. Feature and Topology Probability Estimation
Instead of mapping nodes to fixed vectors, OS-GCL maps them to multidimensional Gaussian distributions $N(\mu, \sigma)$. Simultaneously, the model learns a Bernoulli distribution for the graph topology. The probability of an edge existing is determined by a combination of feature similarity—calculated using Wasserstein distance—and higher-order structural proximity. This allows the model to "look past" superficial observed edges to find more meaningful latent relationships.
2. Probabilistic Message Passing and ProbNCE
OS-GCL introduces a novel graph encoder that propagates these feature distributions over the probabilistic topology, accounting for structural uncertainty during the learning process. To optimize the model, the authors introduced the ProbNCE loss, which contrasts distributions rather than points. This loss can be implemented discretely via reparameterization or continuously by directly minimizing the Wasserstein distance between distributions.
Key Insights and Results
The transition to a probabilistic framework provides a more robust self-supervised signal, particularly in the low-sample regime of one-shot learning.
Performance Benchmarks
OS-GCL was evaluated on seven widely used node classification benchmarks, including citation networks (Cora, CiteSeer, PubMed) and the large-scale ogbn-arXiv dataset.
- Superior Accuracy: OS-GCL consistently outperformed state-of-the-art baselines, showing up to a 3.8% improvement over GRACE and a 2.1% average improvement over methods requiring complex data augmentation.
- Robustness without Augmentation: Unlike many GCL methods that rely on hand-crafted augmentations (such as dropping edges or masking features), OS-GCL achieves high performance purely through its probabilistic modeling.
- Efficiency: Despite the added complexity of modeling distributions, OS-GCL maintains competitive runtimes. By using distribution perturbation in the latent space to create "views," it reduces the need for multiple expensive encoder passes.
Conclusion
By reframing Graph Contrastive Learning as a one-shot problem, OS-GCL highlights the limitations of deterministic representations in sparse data environments. The framework's ability to model uncertainty in both features and topology leads to more distinctive and robust node embeddings, setting a new standard for self-supervised graph learning.
Summary updated