Heterogeneous Graph

Unlike homogeneous graphs, heterogeneous graphs contain multiple node types (e.g., Account, Merchant, Device, IP Address) and multiple edge types (e.g., Transfers_To, Logs_In_From, Owns). Each node and edge type carries its own feature schema, allowing the graph to represent fundamentally different entities within a single interconnected structure. This mirrors real financial ecosystems where accounts, cards, and terminals each have distinct attributes.

In architectures like the Relational Graph Convolutional Network (R-GCN), a distinct weight matrix is learned for each relationship type. This preserves semantic meaning:

• A Transfers_To edge is processed differently from a Shares_Device_With edge
• Prevents information conflation across relationship semantics
• Enables the model to learn that money movement and device sharing carry different fraud signals

Metapaths define composite relationships across node and edge types, such as:

• Account → Transfers_To → Merchant → Charged_Back_By → Account
• Account → Logs_In_From → IP → Logs_In_From → Account

These predefined paths guide the aggregation of non-immediate neighbors, capturing higher-order structural patterns like collusion chains and money laundering layering that span multiple entity types.

Heterogeneous Graph Attention Networks (HAN) apply hierarchical attention at two levels:

1. Node-level attention: Learns the importance of different metapath-based neighbors
2. Semantic-level attention: Learns the importance of different metapaths themselves

This dynamic weighting allows the model to automatically determine that Shares_Device_With relationships may be more fraud-indicative than Transfers_To relationships in certain contexts.

Heterogeneous graphs preserve the native schema of financial data:

• Account nodes carry attributes like avg_balance, account_age_days, kyc_risk_score
• Device nodes carry attributes like os_version, is_rooted, browser_fingerprint
• Transaction edges carry attributes like amount, currency, timestamp, auth_code

This schema preservation eliminates information loss that occurs when flattening diverse entities into a single node type.

Fraud signals propagate across entity boundaries through typed edges. A compromised device can taint multiple accounts that logged in from it. A fraudulent merchant can implicate accounts that transacted with it. The heterogeneous structure enables:

• Transductive inference: Known fraud labels on one entity type inform risk scores on connected entity types
• Multi-hop risk propagation: Risk scores flow through chains like Fraudulent_Account → Device → Clean_Account

A specialized GNN variant designed explicitly for heterogeneous graphs that assigns distinct weight matrices to each relationship type. When processing a financial graph, an R-GCN applies separate transformations for 'transfers_to', 'owns', and 'logged_in_from' edges, preserving the unique semantic meaning of each connection rather than collapsing all interactions into a single homogeneous representation.

A fundamental heterogeneous graph structure consisting of two disjoint node sets—typically account holders and merchants—where edges only connect nodes from different sets. This architecture naturally models directed financial transactions without allowing meaningless account-to-account or merchant-to-merchant edges, enforcing the structural constraints of payment networks.

The process of mapping discrete, heterogeneous nodes—accounts, devices, merchants—into a low-dimensional continuous vector space. In a heterogeneous graph, embeddings must capture not only structural proximity but also type-specific semantics, ensuring that an account node's vector representation encodes its distinct role versus a merchant node, enabling downstream fraud classifiers to reason across entity categories.

The critical feature engineering step of transforming raw relational data into a structured heterogeneous graph. This involves:

• Defining node types: accounts, merchants, IP addresses, devices
• Defining edge types: transfers, logins, ownership, referrals
• Assigning typed attributes: account age, transaction amount, device fingerprint Poor graph construction collapses meaningful heterogeneity, degrading model performance.

The core computational paradigm where nodes in a heterogeneous graph iteratively exchange vectorized information with neighbors. Critically, messages are type-conditioned—an account node aggregates signals differently from 'transferred_to' edges versus 'logged_in_from' edges, using relation-specific transformations to update its hidden state with contextually relevant fraud indicators.

A spatial GNN that introduces self-attention to dynamically weight neighbor importance during aggregation. In heterogeneous financial graphs, GAT mechanisms learn to prioritize suspicious connections—a large transfer from a high-risk jurisdiction receives higher attention weight than a routine micro-transaction—enabling the model to focus on anomalous relational signals without manual feature engineering.