Can assortativity use properties other than degree?

Yes. Networks can be assessed for mixing by categories or other numerical attributes, using suitable definitions.

Does positive assortativity prove that similar nodes prefer one another?

No. It identifies an observed pattern, not the causal mechanism behind it.

Why might software return an undefined result?

A required endpoint variance may be zero, so the correlation's normalization cannot be computed.

AI Glossary / Graphs and symbolic AI / Intermediate

Assortativity

Definition

Assortativity describes how strongly connected nodes in a network resemble one another on a chosen property. Degree assortativity specifically measures whether nodes with many connections tend to connect to other highly connected nodes. Positive values indicate a tendency toward similar degrees; negative values indicate a tendency toward different degrees.

What it means

A network can contain the same number of nodes and links while arranging those links in very different ways. Assortativity helps describe that arrangement by comparing the properties found at opposite ends of connections.

The property must be named. Similarity by degree concerns the number of connections. Similarity by a categorical attribute, such as department, concerns which labels meet across links. These require appropriate definitions and should not be treated as interchangeable measurements.

For degree assortativity, a coefficient near zero means little linear association between the endpoint degrees under the selected calculation. It does not establish that every kind of network structure is absent. Nor does a positive coefficient explain why similar nodes are connected.

How it works

The calculation examines linked endpoint pairs, records the relevant degrees, and measures their correlation. In an undirected network, each link has no preferred direction. A consistent endpoint treatment ensures the result does not depend on how the links happened to be listed.

Directed networks require a choice of degree at each end. An analyst might compare the source node's outgoing degree with the target node's incoming degree. A different pairing can answer a different question and produce a different value.

Weighted networks introduce another choice: a weighted degree can mean the sum of incident edge weights. Analysts should check the software's definition rather than assume weighting simply repeats each observation.

Like other correlations, the coefficient requires variation. If the relevant endpoint degrees have zero variance, the normalized coefficient is undefined. Missing nodes, omitted links, and the selection of a network boundary can also change the result.

Three practical examples

1. A researcher studies whether highly connected accounts mainly connect to other highly connected accounts, while keeping the observed platform boundary explicit.

2. An operations team examines a dependency graph to see whether heavily connected components link to similarly connected components or to many small peripheral components.

3. A collaboration analyst compares connections within and across departments using attribute assortativity, recognizing that this is a different question from degree similarity.

Think of it this way

Think of pairing people according to how many acquaintances each has. Pairings between equally well-connected people resemble positive degree assortativity; pairings between highly connected people and people with few acquaintances resemble negative assortativity.

The analogy describes the pairing pattern only. It does not establish personal preference, influence, or an intentional decision to associate.

Worked example

Consider an undirected star with one center and three leaves. The center has degree 3, and every leaf has degree 1. There are no other links.

Represent each edge in both directions for a symmetric endpoint calculation. The six degree pairs are three copies of (3, 1) and three copies of (1, 3). Both endpoint columns have mean 2.

Every pair has opposite deviations from its column means: either (1, -1) or (-1, 1). The mean product of deviations is -1, and each column's mean squared deviation is 1. Their correlation is therefore -1 divided by the square root of 1 times 1, giving -1.

This small network is perfectly disassortative by degree. That result does not mean its leaves avoid one another deliberately. The coefficient describes the supplied star structure; it does not explain the process that created it.

Illustrative case study

Fictional case: Harbor Systems maps dependencies among internal services. An analyst finds a negative degree-assortativity coefficient and initially calls the architecture poorly organized.

The engineering team examines the graph and sees several shared services connected to many small applications. The coefficient captures this hub-and-periphery pattern, but it does not establish whether the design is appropriate.

The team separates structural description from operational evaluation. It investigates outage impact, redundancy, and ownership using additional evidence, and records which dependency types were included. The fictional lesson is that a network statistic can focus an investigation without settling a business judgment.

Common misconceptions and mistakes

A negative value is not automatically bad, and a positive value is not automatically good. Consequences depend on the network's function and other structural properties.

Zero assortativity is not proof of independence or randomness. It summarizes a particular association, potentially missing nonlinear patterns or community structure.

Degree assortativity is not the same as average degree, clustering, or centrality. Those measurements answer different questions.

Finally, do not replace an undefined value with zero merely to make a dashboard complete. A graph without sufficient degree variation cannot support the same interpretation as a measured zero correlation.

Business use

For business analysis, assortativity is useful as one descriptive feature of a relationship or dependency network. Report the node and edge definitions, direction, weighting convention, observation period, and chosen property alongside the coefficient.

Compare results only when those definitions are sufficiently consistent. If a measurement changes, inspect whether the underlying relationships changed or whether the data collection boundary changed. Combine the statistic with task-specific evidence before making organizational or technical decisions.

Check your understanding

In the three-leaf star above, does a degree-assortativity value of -1 mean that the graph has few edges?

Answer: No. It means that linked endpoint degrees are perfectly negatively correlated in this example. Edge count and degree assortativity are different measurements.

Related terms

Graph
Node degree
Directed graph
Correlation
network
mixing matrix
centrality
clustering coefficient

Learn next

Draw a path, a star, and a regular cycle. List endpoint degrees for each, then identify which examples permit a correlation calculation. Study attribute assortativity afterward to see why changing the measured property changes the question.

Correlation

Sources

Primary sources consulted: NetworkX, degree assortativity coefficient documentation, https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.assortativity.degree_assortativity_coefficient.html ; M. E. J. Newman, Mixing patterns in networks, https://arxiv.org/pdf/cond-mat/0209450 . The star calculation and fictional case are original teaching examples.

Previous: Associative memory Next: Assumed-density filter