Graph Neural Network (GNN)

A spectral-spatial method that generalizes the convolution operation to irregular graph domains. GCNs operate by computing a normalized sum of a node's immediate neighbor features, effectively acting as a low-pass filter that smooths representations across the graph.

• Mechanism: Uses a first-order approximation of spectral graph convolutions.
• Aggregation: Mean-pooling with symmetric normalization via the adjacency matrix.
• Limitation: Assumes equal contribution from all neighbors, lacking an attention mechanism to prioritize important connections.
• Use Case: Semi-supervised node classification on citation networks and social graphs.

Introduces a self-attention mechanism to the message-passing framework, allowing the model to implicitly weigh the importance of a source node's features to a target node. Instead of a fixed mean aggregation, GATs compute dynamic, content-based attention coefficients.

• Mechanism: Multi-head attention over node features, stabilized with LeakyReLU.
• Advantage: Handles heterophilic graphs where dissimilar nodes connect, as the model can learn to ignore noisy neighbors.
• Key Feature: Does not require the full graph structure upfront, enabling inductive learning on completely unseen nodes.
• Use Case: Node classification in protein-protein interaction networks where binding affinity varies.

A spatial-based framework designed for inductive representation learning on large, evolving graphs. GraphSAGE learns a function to generate embeddings by sampling and aggregating features from a node's local neighborhood, rather than training individual embeddings for each node.

• Mechanism: Samples a fixed-size set of neighbors and applies a differentiable aggregation function (mean, LSTM, or max-pooling).
• Scalability: Enables batched training on massive graphs that cannot fit into GPU memory.
• Key Feature: Generalizes to unseen nodes immediately after training, critical for dynamic platforms with new users or content.
• Use Case: Generating embeddings for new Reddit posts or newly published academic papers.

A general, unifying framework that formalizes the commonalities of most spatial GNNs. MPNNs abstract the process into two distinct phases: a message function that computes information passed along edges, and an update function that uses aggregated messages to compute new node states.

• Phases: Message Passing Phase (propagation) and Readout Phase (global pooling).
• Flexibility: Edge features (like bond type in molecules) are first-class citizens in the message computation.
• Foundation: Provides the theoretical blueprint for more complex architectures used in computational chemistry.
• Use Case: Predicting quantum mechanical properties of molecules for drug discovery.

A theoretically grounded architecture designed to maximize the discriminative power of GNNs. GIN proves that the sum aggregation function is injective, making the model as powerful as the Weisfeiler-Lehman (WL) graph isomorphism test for distinguishing different graph structures.

• Mechanism: Uses a multi-layer perceptron (MLP) to process the sum of a node's features and its neighbors' features.
• Advantage: Captures structural information that mean or max aggregators might lose, such as the multiplicity of identical substructures.
• Key Feature: A learnable parameter epsilon controls the weight of the central node versus its neighbors.
• Use Case: Graph classification tasks requiring structural fingerprinting, like social network analysis.

A variant of the GraphSAGE framework that employs a standard LSTM architecture as the aggregation function. While LSTMs are inherently sequential, applying them to a random permutation of a node's neighbors provides a powerful, expressive aggregator that can capture complex dependency patterns.

• Mechanism: Feeds a randomly shuffled sequence of neighbor feature vectors into an LSTM cell.
• Advantage: Higher representational capacity than simple mean or max pooling.
• Caveat: Not permutation-invariant by design, but the random shuffling acts as a practical regularization technique.
• Use Case: Scenarios where the order of neighbors might carry latent semantic meaning, such as temporal citation networks.

A foundational spatial-based GNN that generalizes the convolution operation to irregular graph domains. GCNs operate by computing a node's new representation as a weighted average of its own features and those of its immediate neighbors, followed by a non-linear activation. This layer-wise propagation rule, derived from a first-order approximation of spectral graph convolutions, enables efficient semi-supervised node classification by diffusing label information across the graph structure.

An architecture that introduces a self-attention mechanism into the message-passing framework. Instead of treating all neighbors equally, GATs compute dynamic, learnable attention coefficients for each edge, allowing nodes to implicitly weigh the importance of different neighbors during aggregation. This enables the model to focus on the most relevant parts of the graph and handle heterophilic neighborhoods where connected nodes may have dissimilar features.

A unified framework that formalizes the common operations of most spatial GNNs into two core functions: a message function that computes information to be sent along edges, and an update function that aggregates incoming messages to produce a new node state. This abstraction generalizes GCNs, GATs, and GraphSAGE, providing a consistent mathematical language for designing and analyzing graph learning models.

An inductive framework that generates node embeddings by sampling and aggregating features from a node's local neighborhood. Unlike transductive methods that require the full graph during training, GraphSAGE learns a set of aggregator functions (mean, LSTM, pooling) that can be applied to previously unseen nodes. This makes it suitable for dynamic, evolving graphs where new entities are continuously added.

The process of mapping discrete graph nodes to low-dimensional, dense vector representations that preserve structural and relational properties. Classical methods like DeepWalk and node2vec use random walks to capture neighborhood topology, while GNN-based embeddings encode both graph structure and node features. These vectors serve as input for downstream tasks such as node classification, link prediction, and clustering.

A theoretically grounded GNN architecture designed to be as powerful as the Weisfeiler-Lehman (WL) graph isomorphism test in distinguishing different graph structures. GIN achieves this by using a sum aggregation over neighbor features combined with a multi-layer perceptron, ensuring that the model can capture and differentiate complex structural patterns that simpler aggregators like mean or max might conflate.