Traditional Node Features in Graph ML | CS224W Lecture 2.1

Added:

End-to-End Representation Learning and Graph Neural Networks (GNNs)

While traditional node-level feature engineering relies on hand-crafted, domain-specific metrics like centrality and graphlets, the modern paradigm favors representation learning. Critics of traditional features point out several key limitations: they require substantial domain expertise, are computationally expensive to calculate (e.g., subgraph/graphlet counting is highly intractable on large scales), and are static, meaning they cannot adapt to the specific downstream task. In contrast, Graph Neural Networks (GNNs) and node embedding methods automatically learn low-dimensional, task-optimized vector representations directly from the graph topology and raw node attributes. This end-to-end learning approach generalizes better to unseen data, scales more effectively to massive networks, and captures complex, non-linear structural patterns that manual feature engineering often misses.

Basic graph theory concepts, including nodes, edges, adjacency matrices, and neighborhood definitions.
Fundamentals of machine learning feature engineering, specifically how structured data is represented as feature vectors for model input.
Elementary linear algebra, such as vector spaces and matrix operations, to understand graph representations.
Understanding of basic graph traversal algorithms and concepts like paths, walks, and degrees.
Traditional link-level and graph-level feature engineering techniques (e.g., Katz index, common neighbors, and graph kernels).
Introduction to node embedding methods (representation learning) such as DeepWalk and Node2Vec that automate feature extraction.
Graph Neural Networks (GNNs) like GCNs and GraphSAGE, which combine node features and graph structure end-to-end.
Practical applications of node features in downstream machine learning tasks like node classification and community detection.
149.7K views1.5Klikes27:30@stanfordonlineOriginal Release: 2021-04-15

In graph machine learning, node-level features are categorized into importance-based features (like node degree, eigenvector centrality, betweenness centrality, and closeness centrality) that capture a node's significance in the network, and structure-based features (like clustering coefficient and graphlet degree vector) that describe the local topological patterns around a node; these features enable machine learning models to make predictions about individual nodes by encoding both their intrinsic properties and their position within the broader network structure.