KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Heterogeneous Graph Neural Network
Neural network architecture specifically designed to process heterogeneous graphs containing multiple types of nodes and edges with different semantic relationships. These models capture the structural and semantic complexity of multi-relational networks while preserving information specific to each type.
Metapath
Sequence of node types and relationships defining a specific semantic path in a heterogeneous graph, used to guide information aggregation and capture complex relational patterns. Metapaths allow encoding high-level relationships between different types of entities in the network.
Type-aware Aggregation
Aggregation mechanism that differentiates and separately processes information from different types of neighboring nodes using type-specific aggregation functions. This approach preserves the inherent semantics of each entity category during information propagation.
Heterogeneous Attention Mechanism
Variant of the attention mechanism adapted to heterogeneous graphs that calculates attention weights considering both node types and relationship types. This mechanism allows modeling the relative importance of different neighbors based on their semantic nature.
Relation-specific Transformation
Linear or nonlinear transformation operation applied to node features according to the type of relationship connecting them, allowing capture of different relational semantics. Each edge type has its own independently parameterized transformation matrix.
Cross-type Message Passing
Process of information propagation between different types of nodes in a heterogeneous graph, using communication schemes adapted to each source-target type combination. This mechanism facilitates semantically coherent information transfer between heterogeneous entities.
Node Type Embedding
Learned vector representation for each node type in a heterogeneous graph, integrated into the final node embeddings to preserve categorical information. These embeddings are concatenated or added to structural embeddings to enrich the representation.
Semantic-level Propagation
Propagation strategy that organizes node updates according to semantic levels defined by relationship types and graph hierarchy. This approach ensures that information flows coherently with the semantic structure of the network.
Metapath2Vec
Representation learning algorithm for heterogeneous graphs based on metapath-guided random walks, generating embeddings that preserve semantic and structural proximity between different types of nodes. It extends the Word2Vec paradigm to multi-relational networks.
HAN (Heterogeneous Graph Attention Network)
HGNN architecture using a hierarchical attention mechanism with node-level and metapath-level attention to automatically capture the importance of different semantic paths. HAN learns to weight relevant metapaths for each specific task.
RGCN (Relational Graph Convolutional Network)
Extension of GCNs to multi-relational graphs using transformations specific to each relation type and regularization techniques to handle the large number of parameters. RGCN is particularly effective for knowledge bases and graphs with many relations.
HGT (Heterogeneous Graph Transformer)
Transformer architecture adapted for heterogeneous graphs using attention mechanisms specific to node and relation types, with shared and type-specific parameters. HGT efficiently captures heterogeneous dependencies and scales to large graphs.
MAGNN (Metapath Aggregated Graph Neural Network)
HGNN model that aggregates information according to metapath instances using metapath encoders and intra-metapath attention mechanisms. MAGNN captures complex semantics of multi-hop relations in heterogeneous graphs.
Type-constrained Random Walk
Random walk algorithm on heterogeneous graphs where transitions are constrained by allowed node and relation types according to predefined schemas. This approach generates semantically coherent sequences for representation learning.
Heterogeneous Information Network
Formal graph structure containing objects of multiple types and semantically diverse relations, serving as the basis for HGNNs. HINs model complex systems where different entities interact according to varied relational schemas.
Metapath-based Sampling
Neighbor sampling strategy in HGNNs that selects nodes based on their participation in relevant metapaths, optimizing computational efficiency while preserving relational semantics. This technique reduces complexity while maintaining embedding quality.