Can two links with equal embeddedness have different dispersion?

Yes. Their common neighbors can have different connections among themselves.

Is default normalized dispersion a probability?

No. It is a normalized structural score.

Does high dispersion prove a particular relationship type?

No. Such an interpretation requires evidence beyond the graph feature and depends on the setting.

AI Glossary / Graphs and symbolic AI / Intermediate

Dispersion

Definition

Dispersion, in the network-analysis sense used here, measures how separated the mutual neighbors of two connected nodes are from one another. A connection can have high dispersion when its shared neighbors occupy different parts of the surrounding network rather than one tightly connected group. The supplied glossary term refers to the NetworkX measure based on Backstrom and Kleinberg's work.

What it means

Counting mutual neighbors measures embeddedness: how many shared connections a pair has. Dispersion asks an additional question about the relationships among those shared connections. Two pairs can have the same number of mutual neighbors yet very different dispersion because one shared neighborhood is cohesive and the other is fragmented.
This is a structural graph measure, not the statistical variance of numerical values or physical spreading through space. It also does not directly reveal the meaning, quality, or strength of a real relationship. Interpreting it requires knowing what nodes and edges represent and how completely they were observed.

How it works

For an undirected ego-network interpretation centered on u, identify the common neighbors of u and a connected node v. Examine unordered pairs of these common neighbors. In the NetworkX implementation, a pair contributes to the raw score when its members are not directly connected and have no shared neighbor, other than the excluded endpoints u and v, within the relevant ego neighborhood.
The raw score counts such separated pairs. With the documented default normalization, it is divided by embeddedness when that denominator is nonzero. The function also exposes parameters that can change the normalization. Consequently, raw and normalized values should not be mixed, and the graph neighborhood used for the calculation must be stated.

Three practical examples

A synthetic social-network exercise compares links whose mutual contacts form one group with links whose mutual contacts span disconnected groups.
A collaboration-network analysis examines whether a connection shares contacts across otherwise separated project circles.
A graph-methods benchmark varies connections among mutual neighbors to test whether an implementation captures separation rather than only counting shared contacts.

Think of it this way

Think of two organizers who know the same three volunteers. If the volunteers all work together, the shared contacts form one circle. If the volunteers belong to separate circles, the organizers' shared connection reaches across a more dispersed neighborhood. The analogy is limited because the score uses recorded edges and specific graph rules, not a complete understanding of people's social lives.

Worked example

Construct an undirected graph with nodes u, v, a, b, and c. Include the edge u - v, and connect each of a, b, and c to both u and v. Add no other edges. The common-neighbor set is {a,b,c}, so embeddedness is 3. Its unordered pairs are {a,b}, {a,c}, and {b,c}.
After excluding u and v as connecting intermediaries, none of those pairs is directly connected or has another shared neighbor in this graph. Each contributes 1, giving raw dispersion 3. Default normalized dispersion is 3/3 = 1.
Now add all three edges a - b, a - c, and b - c. Embeddedness remains 3, but every common-neighbor pair is directly connected, so raw and default normalized dispersion both become 0. The change isolates neighborhood structure while keeping the number of mutual neighbors fixed. The normalized value 1 in the first example is a score, not a probability or a universal maximum.

Illustrative case study

At fictional Maple Collaboration, an analyst studies a synthetic project network to understand a graph feature before applying it to approved organizational data. Two links have equal mutual-contact counts but different patterns among those contacts. The analyst uses dispersion to describe that structural difference.
The report avoids assigning personal relationship labels from the feature alone and documents which interactions the graph records. This fictional case illustrates interpretation and scope. It does not claim that dispersion identifies a person's role, private relationships, performance, or intentions with certainty.

Common misconceptions and mistakes

Dispersion is not merely the number of mutual neighbors. It examines connections among them. A normalized score is also not automatically bounded by 1: dividing a pair count by a node count is different from converting the result into a probability.
Do not silently treat a directed or weighted relationship model as identical to the simple undirected setting. Missing edges can make contacts appear separated when they are actually connected. The original research explored a specific social-network question, but that does not justify transferring its findings unchanged to every network or using one score as proof of a private personal attribute.

Business use

Businesses can use dispersion as an exploratory feature when the graph and use case support a meaningful interpretation. Compare it with embeddedness and inspect representative local structures rather than relying on a ranking alone. Define which interactions count as edges, the time window, and the normalization parameters. For people-related networks, keep conclusions tied to observed network structure and an appropriate authorized purpose; do not turn an ambiguous structural signal into an unsupported personal claim.

Check your understanding

Why does adding the triangle among a, b, and c reduce the worked score without changing embeddedness?

Answer: The same three nodes remain mutual neighbors, but their pairs are now directly connected and no longer contribute to raw dispersion.

Related terms

Ego network
embeddedness
common neighbor
social network analysis
graph feature
tie strength
local clustering
network topology

Learn next

Learn embeddedness and clustering next. Construct two graphs with the same mutual-neighbor count but different local connections, then calculate raw and normalized dispersion separately. Explain the structural difference without assigning an unobserved social meaning.

Sources

Primary sources: NetworkX official dispersion documentation, https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.centrality.dispersion.html ; implementation, https://networkx.org/documentation/stable/_modules/networkx/algorithms/centrality/dispersion.html ; Backstrom and Kleinberg, original research, https://arxiv.org/abs/1310.6753 . The graph and fictional case are original illustrations.

Previous: Disparity map Next: Dispersive flies optimization