DOMINANT

DOMINANT uses a shared Graph Convolutional Network (GCN) encoder to compress both node features and the adjacency matrix into a latent space. Two separate decoders then reconstruct the node attributes and the graph structure independently. This dual-decoder design forces the latent representation to capture the complex interplay between what a node is (its features) and how it connects (its topology), enabling the detection of anomalies that are subtle in one view but obvious in the other.

Anomalies are ranked by a combined score: (1 - α) * Structure_Error + α * Attribute_Error, where α is a hyperparameter balancing the two perspectives. A node is flagged as anomalous if it cannot be accurately reconstructed by the model. For example, a fraudulent account might have a high structural error because it connects to many disparate, low-credit nodes, and a high attribute error if its transaction frequency is an outlier for its stated profile. This unified score provides a principled ranking mechanism.

The architecture consists of three core components:

• Encoder: A multi-layer GCN that maps the input graph (features X and adjacency A) to low-dimensional node embeddings Z.
• Structure Decoder: Reconstructs the adjacency matrix  from the inner product of latent embeddings ( = sigmoid(ZZᵀ)), predicting link existence.
• Attribute Decoder: Another GCN that reconstructs the original node features X̂ from the latent embeddings Z. This design is fully unsupervised, requiring no labeled fraud data for training.

Unlike traditional GNNs that assume connected nodes are similar (homophily), fraud graphs often exhibit heterophily, where fraudsters connect to many legitimate accounts to appear normal. DOMINANT's dual reconstruction mechanism is inherently robust to this. A fraud node connecting to benign nodes will have a latent representation that is a poor compromise, leading to high reconstruction error in either its attributes or its predicted links, thus exposing the anomaly despite the deceptive connections.

In the original paper, DOMINANT was evaluated on real-world datasets like Amazon and YelpChi for opinion spam detection. It consistently outperformed baseline methods like LOF, SCAN, and standard autoencoders by a significant margin in terms of ROC-AUC. The joint learning approach proved particularly effective at catching camouflaged anomalies—nodes that manipulate their features or connections to evade detection by simpler, single-view methods.

The core computation involves GCN layers with a complexity linear in the number of edges, O(|E|). The structure decoder's inner product (ZZᵀ) is O(N²) in theory but is optimized using negative sampling during training, where only a small subset of non-edges are reconstructed. This makes DOMINANT scalable to large graphs with millions of edges, suitable for real-world financial transaction networks where full adjacency matrix reconstruction is computationally prohibitive.