What is node classification?

Predicting labels for nodes in a network using both their features and their connections to other nodes.

What assumption does it rely on?

Homophily: connected nodes tend to share labels, which holds for fraud rings, communities and contagion, and fails at mixing boundaries.

Why is evaluation special on graphs?

Random splits leak: a test node's neighbors may be in training, so evaluation must separate communities or cut by time.

What are simple starting methods?

Label spreading and neighbor-vote baselines, before graph neural networks earn their complexity.

Do I need a graph neural network?

Not to start: label spreading and neighbor-vote baselines show whether the connections carry signal before the heavier machinery is justified.

AI Glossary / AI and machine learning / Advanced

Node Classification

Definition

Node classification is the task of labeling the nodes of a network: given a graph of connected items and labels for some of them, predict the labels of the rest, using both their own features and their connections.

What it means

The setting appears whenever relationships carry information: customers connected by transactions, accounts by shared devices, papers by citations. A node's neighborhood whispers its label.

The core assumption is homophily: connected things resemble each other. Fraud rings, interest communities and churn contagion all show it, and the assumption's strength in a given graph decides how much the connections add.

Methods range from simple label spreading, where labels diffuse across edges, to graph neural networks that learn how each node should combine its own features with its neighbors' information.

The evaluation needs care: random splits leak, because a test node may sit beside its training neighbors, so honest testing separates connected groups or predicts forward in time.

For a business, node classification is how relationship data becomes predictions: the network of who connects to whom is often the strongest signal a company owns and the least used.

The task sits inside graph machine learning as its most common job, and success stories from fraud to recommendation are usually node classification in work clothes.

How it works

Build the graph deliberately: what the nodes are, what an edge means, and why connected nodes should share labels.

Start with the baseline of node features alone, so the graph's actual contribution is measured rather than assumed.

Apply label spreading or a graph model that aggregates neighborhood information into each node's representation.

Evaluate with leakage-aware splits: separated communities or time-based cuts, never random node splits on a connected graph.

Inspect errors by neighborhood, since failures concentrate where the homophily assumption breaks, at the boundaries between groups.

Three practical examples

A bank flagging accounts likely to be fraudulent from their transaction connections, when only a fraction are confirmed.
A social platform predicting professional interests for unlabeled members from their connections' declared interests.
A telco predicting churn risk spreading through call networks, where one departure raises its neighbors' risk measurably.

Think of it this way

Think of guessing which candidate a street supports from the posters in a few windows: the street's other windows are strong evidence, because neighbors, in politics as in graphs, tend to agree.

Worked example

the ring that the network saw

Suppose a payment platform labels only confirmed fraud, two per cent of accounts. The figures are illustrative.

Features alone catch the obvious cases, but node classification across the transaction graph flags a cluster of forty unlabeled accounts wired tightly to confirmed ones.

Review confirms thirty-one are coordinated: individually unremarkable, collectively damning, and invisible to any account-by-account model.

The graph's contribution is the pattern across accounts: no single node's features held the finding, and only the connections expressed it.

Illustrative case study

A fictional marketplace trained a node classifier on its seller network and evaluated it with random splits, reporting 95 per cent accuracy.

In production the model scored near chance, because the random split had hidden the truth: test nodes had sat among their training neighbors, and the real task, labeling genuinely new territory, was never tested.

The invented lesson: on a graph, a random split is a leak. Evaluate across separated communities or forward in time, or the score is measuring memory, not skill.

The repair was structural: split the graph by community before splitting by node, and the honest score, lower but real, finally matched production.

Common misconceptions and mistakes

Graph data is just more features. Edges are a different kind of evidence with their own assumptions and their own leakage traps.
Connections always help. They help under homophily; when connected nodes differ systematically, the same edges mislead.
Random test splits are fine. On connected data they leak labels through edges, and honest evaluation must cut the graph deliberately.

Business use

Inventory your relationship data as an asset: transactions, shared devices, referrals and calls are prediction fuel.

Establish the feature-only baseline before buying graph machinery, so the network's lift is priced honestly.

Demand leakage-aware evaluation from any graph-model vendor: separated communities or time cuts, in writing.

Respect the boundaries: where groups mix freely, neighborhood evidence weakens, and errors will concentrate there.

Document edge definitions in the model record, since what counts as a connection is a business decision that shapes every prediction.

Check your understanding

Why does a random train-test split inflate a node classifier's measured accuracy?

Answer: Because on a connected graph a test node's neighbors carry their labels, and if those neighbors were in the training set the model has effectively seen the answer, so honest evaluation separates connected groups or predicts forward in time.

Related terms

Node (decision tree)
Graph neural network
Classification
label spreading
network
homophily

Learn next

AI for Business Owners helps you judge whether relationship data is being used or wasted: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S9] NetworkX documentation; [S11] Murphy, Probabilistic Machine Learning: An Introduction. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Node Next: Node (decision tree)