AI-woordenlijst
Het complete woordenboek van kunstmatige intelligentie
Graph Convolutional Network (GCN)
Neural network architecture that applies convolution operations on graph data structures by aggregating features from neighboring nodes to learn node representations.
Neighborhood Aggregation
Process of combining features from neighboring nodes to update a target node's representation, typically through mean, sum, or max operations.
Spectral Methods
Approach based on graph spectral theory using eigenvalue decomposition of the Laplacian to define convolution operations on graphs.
Spatial Methods
Direct approach applying convolution operations in node space by physically aggregating neighbor features without spectral transformation.
Graph Laplacian
Matrix representing the structure of a graph, defined as the difference between the degree matrix and the adjacency matrix, fundamental for spectral methods.
Feature Propagation
Mechanism by which node features propagate through the graph via successive convolution layers, capturing neighborhood information at different scales.
Semi-Supervised Learning
Learning paradigm where GCNs use both labeled and unlabeled data to improve classification performance on graphs.
Node Classification
Fundamental task where GCNs predict node labels using the graph structure and features of neighboring nodes.
Link Prediction
Application of GCNs to predict the existence of links between node pairs by learning representations that capture connection probability.
Graph Classification
Classification task at the whole graph level where a GCN learns a global representation of the graph to predict a label for the entire structure.
Graph Attention Network (GAT)
Variant of GCNs incorporating attention mechanisms to dynamically compute edge importance weights during feature aggregation.
Over-smoothing
Phenomenon where node representations become indistinguishable after multiple convolution layers, losing their individual discriminability.
Graph Sampling
Technique for sampling subgraphs or neighborhoods to efficiently train GCNs on large-scale graphs.
Graph Pooling
Hierarchical reduction operation that combines or eliminates nodes to create coarser graph representations for graph-level classification.
Heterogeneous GCN
Extension of GCNs designed to handle graphs containing multiple types of nodes and/or edges with type-specific aggregation mechanisms.
Temporal GCN
Variant of GCNs that captures the dynamic evolution of graphs over time by integrating recurrent or temporal mechanisms with graph convolution.