Contrastive Learning on Graphs

The core objective is to pull semantically similar graph views together in the embedding space while pushing dissimilar ones apart. This is achieved by maximizing the mutual information between a local node representation and a global graph summary vector, or between two augmented views of the same subgraph. The model learns to identify what makes a node or graph unique, capturing structural and feature-based patterns that are invariant to stochastic augmentations.

Unlike computer vision, augmentations must preserve the underlying semantics of the graph. Common strategies include:

• Edge Dropping: Randomly removing edges to simulate incomplete transaction data.
• Node Feature Masking: Randomly zeroing out account attributes to force the encoder to rely on structural context.
• Subgraph Sampling: Extracting localized neighborhoods via random walks to create positive pairs. These operations generate diverse views of the same financial entity, teaching the model to be invariant to noise.

A discriminator network scores the agreement between local node embeddings and a global readout summary of the entire graph. The positive sample is the pairing of a node with its native graph's summary, while negative samples pair the same node with summaries from alternative graphs in the batch. This forces the encoder to capture globally relevant structural roles, such as identifying a money mule node that bridges otherwise disconnected clusters.

This approach contrasts different augmented views of the same node against views of other nodes in the batch. For a given account, one augmented view (e.g., with dropped edges) is treated as the anchor, and another augmented view of the same account is the positive sample. All other accounts in the batch serve as negative samples. This teaches the encoder to produce consistent representations for the same entity despite structural perturbations, a critical property for identity resolution.

Contrastive learning serves as a powerful pretext task for fraud detection where labels are scarce. The pre-trained graph encoder generates rich, structural node embeddings that can be frozen and fed into a simple downstream classifier (e.g., logistic regression) trained on a very small set of known fraud labels. This semi-supervised pipeline significantly outperforms end-to-end supervised GNNs when labeled anomalies are rare, as the encoder has already learned the topology of normal behavior.

The quality of learned representations depends heavily on the difficulty of negative samples. Strategies include:

• Random Batch Negatives: Treating other nodes in the mini-batch as negatives.
• Hard Negative Mining: Actively selecting nodes that are structurally similar but belong to different graphs or classes.
• Adaptive Negative Sampling: Dynamically adjusting the sampling distribution to focus on nodes the discriminator currently finds confusing. In financial graphs, hard negatives prevent the model from collapsing representations of distinct but topologically similar accounts.

The core objective function in contrastive learning. The model is trained to maximize the mutual information between a local node representation and a global graph summary vector.

• Positive Pairs: A node embedding and the graph summary from the same graph.
• Negative Pairs: The same node embedding paired with summaries from alternative graphs in the batch.
• Discriminator Network: A neural network that scores agreement between local-global pairs, forcing the encoder to capture globally relevant structural patterns.

The mechanism for generating distinct views of the same graph without semantic alteration. Effective augmentations are critical for preventing representational collapse.

• Edge Dropping: Randomly removing a subset of edges to create a sparser view, simulating missing transaction data.
• Node Feature Masking: Zeroing out a random fraction of node attributes, forcing the encoder to rely on structural context.
• Subgraph Sampling: Extracting a contiguous subgraph via random walk to create a localized view.
• Diffusion: Generating a new adjacency matrix based on graph diffusion kernels like Personalized PageRank.

The technique for selecting non-matching pairs to contrast against. The choice of negative samples directly impacts the quality of the learned representations.

• Intra-View Negatives: Other nodes within the same augmented graph view are treated as negatives to the anchor node.
• Inter-View Negatives: Nodes from a different graph in the training batch serve as negatives.
• Hard Negative Mining: Intentionally selecting nodes that are structurally similar but belong to different graphs to force the encoder to learn more discriminative features, preventing the model from solving the task too easily.

A seminal framework for Graph Contrastive Learning that formalizes the augmentation pipeline. It applies stochastic transformations to an input graph to generate correlated views.

• Process: Apply two augmentations (e.g., node dropping and edge perturbation) to create views G_i and G_j.
• Encoding: A shared GNN encoder generates graph-level representations for both views.
• Projection Head: A non-linear MLP maps representations to a latent space where the contrastive loss (NT-Xent) is applied.
• Utility: Pre-trains a GNN backbone that can be fine-tuned on a small set of labeled fraud transactions.

A foundational architecture for unsupervised graph learning that contrasts local node patches against a global summary of the entire graph.

• Encoder: A GCN generates patch representations h_i for each node.
• Readout: A permutation-invariant function averages all node embeddings into a global summary vector s.
• Discriminator: Scores the summary-vector pair (h_i, s) as real, while pairing h_i with a corrupted summary s̃ from a shuffled feature graph as fake.
• Fraud Application: Learns to identify nodes whose local structure is anomalous compared to the global transaction network norm.

An adaptation of Bootstrap Your Own Latent that eliminates the need for negative samples, a major bottleneck in contrastive learning.

• Asymmetric Architecture: Uses an online encoder and a target encoder, where the target network's weights are an exponential moving average of the online network.
• Predictor: A shallow MLP maps the online representation to predict the target representation of the same node under a different augmentation.
• Collapse Prevention: Avoids representational collapse without negatives by using a stop-gradient operation on the target branch, making it highly scalable for massive financial graphs.