Graph Structure Learning

Unlike standard GNNs that treat the input graph as fixed ground truth, GSL treats the adjacency matrix as a learnable parameter. The model simultaneously updates edge weights or discrete connections alongside neural network weights, optimizing the graph structure to be maximally informative for the downstream fraud detection task. This bi-level optimization directly minimizes the final classification loss with respect to both the graph topology and the GNN parameters.

GSL acts as an adaptive denoising mechanism for noisy transaction graphs. It learns to prune spurious or irrelevant edges—such as low-value noise transactions—while strengthening or adding edges that represent genuine behavioral similarities. This process is critical in fraud scenarios where adversaries deliberately inject misleading connections to camouflage their activity, effectively reconstructing a cleaner, more discriminative graph for message passing.

Many GSL frameworks use deep metric learning to infer a graph from raw node features when an explicit input graph is unavailable or unreliable. The model learns a task-specific similarity kernel—such as cosine or Mahalanobis distance—in the embedding space. Nodes that are semantically close in this learned latent space are connected, creating an entirely new graph structure optimized for the specific anomaly detection objective.

Fraudsters actively manipulate relational structures through techniques like feature camouflage and relation perturbation. GSL provides inherent robustness by continuously re-evaluating and rewiring the graph during training. If an adversary adds fake connections to legitimate nodes to appear normal, the structure learning module can learn to down-weight or sever those edges, isolating the malicious entity and preventing the adversarial signal from propagating through the GNN.

To prevent the learned graph from collapsing into a trivial solution—such as a complete graph or an empty one—GSL frameworks incorporate structural priors as regularization terms. Common constraints include:

• Sparsity regularization: Encouraging a low number of edges per node
• Smoothness regularization: Promoting connections between nodes with similar features
• Community preservation: Maintaining the original graph's macro-level cluster structure These priors ensure the learned topology remains physically meaningful and computationally efficient.

Learning a discrete graph structure is a non-differentiable combinatorial problem. GSL addresses this through continuous relaxations and iterative projection techniques. The model learns a continuous probabilistic adjacency matrix during gradient descent, then applies techniques like the Gumbel-Softmax reparameterization or projected gradient descent to sample or project the soft connections back into a valid discrete graph structure for the forward pass of the GNN.

The foundational feature engineering process that transforms raw relational data into a structured graph format. This involves defining node types (accounts, merchants, devices), edge types (transfers, logins, ownership), and feature vectors (transaction amounts, timestamps, risk scores). The quality of graph construction directly determines the signal available for downstream structure learning, as poorly defined edges introduce noise that must later be denoised.

A class of neural networks designed to perform inference on graph-structured data. GNNs learn node representations by recursively aggregating information from local neighborhoods through message passing. In the context of structure learning, the GNN serves as the downstream classifier whose performance is optimized jointly with the graph topology, creating a feedback loop where the model learns which connections are predictive of fraud.

The discipline of hardening fraud detection models against deliberate evasion attacks. Fraudsters actively manipulate transaction patterns to evade detection, which can be modeled as adversarial perturbations to the graph structure. Graph structure learning acts as a defense mechanism by identifying and removing these adversarially injected edges, restoring the clean graph topology before classification occurs.

The task of identifying nodes, edges, or subgraphs whose structural patterns deviate significantly from the majority. Structure learning complements anomaly detection by first denoising the reference graph, ensuring that anomalies are measured against a clean baseline rather than a corrupted one. Frameworks like DOMINANT jointly learn attribute and structural patterns using GCN-based autoencoders, ranking entities by reconstruction error.

A self-supervised paradigm that trains graph encoders to maximize mutual information between different augmented views of the same graph. Applied to structure learning, contrastive objectives can learn robust representations without labeled fraud data by treating the original graph and a corrupted version as positive and negative pairs. This forces the encoder to distinguish genuine structural patterns from injected noise.

A graph learning task focused on predicting the likelihood of future or missing connections between nodes. Structure learning leverages link prediction as an auxiliary objective to assess edge plausibility: edges with low predicted probability under a learned model are candidates for removal, while high-probability missing edges may represent undiscovered relationships that should be added to the graph.