Does it require model training?

The basic index is calculated directly from a graph; it does not fit parameters from labeled examples.

Can it score directed networks?

Variants are possible, but the neighborhood and degree definitions must be specified. The standard NetworkX implementation is for undirected graphs.

Is it bounded by one?

No. Its summed contributions can exceed one.

AI Glossary / Data science and analytics / Advanced

Adamic - Adar index

Definition

The Adamic - Adar index is a graph similarity score that adds up contributions from the neighbors shared by two nodes, giving more weight to shared neighbors with fewer connections. It is often used to rank possible missing or future links. The score is a structural clue, not a probability that two people or objects should be connected.

What it means

Two people sharing a colleague who knows only a few people can provide a different signal from sharing a colleague connected to almost everyone. Adamic - Adar expresses this intuition through inverse logarithmic weighting. A widely connected common neighbor still contributes, but contributes less.
The interpretation depends on what the graph represents. A connection might mean collaboration, an approved referral or a link between documents. The same numerical score cannot establish trust, compatibility or business value across all these meanings. Before using the index, define the nodes, the edges and the decision the ranking will inform.

How it works

For distinct nodes u and v in a simple undirected graph, find their common neighbors. For each common neighbor w, calculate 1 / ln(degree(w)), then add these values. Degree is the number of neighbors of w in the graph, not merely its connections to u and v. With no common neighbors, the sum is zero.
Using natural logarithms fixes the numerical convention. A common neighbor of two distinct nodes has degree at least two in this graph setting, so its denominator is positive. Self-loops, directed edges and repeated edges require explicit treatment rather than an unexamined reuse of the formula.
Rank eligible candidate pairs by their scores. Evaluate the ranking against an appropriate outcome; the calculation itself does not learn what a successful recommendation means.

Three practical examples

A professional community ranks potential introductions using shared collaborators. A specialist with a small collaboration circle contributes more than a highly connected conference organizer. Members still choose whether to accept an introduction.
An editorial team looks for missing links between related articles in a curated, undirected relationship graph. An editor checks whether each proposed link helps readers before publishing it.
A research team uses the score as a baseline for predicting future collaborations from an earlier network snapshot. It compares performance with a simple common-neighbor count before adopting a more complex model.

Think of it this way

Consider recommendations from mutual acquaintances. Knowing the same person who introduces almost everyone is a broad signal. Knowing the same person with a small, focused circle may be more distinctive. Adamic - Adar formalizes that difference in connectedness. It does not establish that the selective acquaintance actually recommends either person, or knows them well.

Worked example

Suppose candidate nodes A and B share exactly two neighbors, X and Y. X has degree 2 and Y has degree 10 in the full graph. Using natural logarithms, their contributions are 1 / ln(2), approximately 1.443, and 1 / ln(10), approximately 0.434. The score for A and B is therefore approximately 1.877.
Another candidate pair, C and D, shares exactly two neighbors, each with degree 10. Its score is approximately 0.434 + 0.434 = 0.869. A common-neighbor count gives both pairs a score of 2, while Adamic - Adar ranks A - B higher because one shared neighbor is less widely connected.
The score 1.877 is not a 187.7% probability. These are invented graph values demonstrating a ranking rule. Neither pair has been shown to form a useful connection, and a higher score does not establish causation.

Illustrative case study

In a fictional project, Cedar Exchange wants to suggest collaborators within an opt-in professional community. Its existing system counts shared collaborators and often produces ties. The team adds Adamic - Adar as a comparison, using only collaboration records available before each evaluation date.
The weighted ranking changes some suggestions, but reviewers notice that newcomers with few recorded links receive limited exposure. Cedar keeps a separate newcomer discovery route and allows members to decline recommendations. It evaluates accepted, useful introductions alongside coverage rather than celebrating changes in the score alone.
The pilot produces a more informative comparison between two simple ranking methods. It does not prove that low-degree intermediaries are better judges of people. The team records that missing collaboration data and the platform's membership patterns influence the graph and its recommendations.

Common misconceptions and mistakes

Adamic - Adar does not simply count common neighbors, and it does not divide by the degree of each candidate node. The weighting uses each shared neighbor's degree. It also differs from resource allocation scoring, which uses inverse degree rather than inverse log degree.
A zero score means no shared neighbors under the chosen graph representation; it does not mean a connection is impossible or undesirable. Testing with future edges already present in the input graph leaks information into the evaluation.
Treating popularity reduction as fairness is another mistake. The weighting changes one structural influence, while unequal participation and missing records can still shape the result.

Business use

Use the index as an interpretable baseline where shared connections plausibly inform discovery. Specify candidate eligibility before ranking: existing links, unavailable participants or unsuitable document pairs may need exclusion.
Measure useful outcomes, coverage and review effort. Compare against simpler alternatives using the same candidate set and time boundary. A gain on a narrow sample may disappear when applied to the wider population. Retain the contributing neighbors so an authorized reviewer can understand a suggestion, while avoiding unnecessary disclosure of personal relationship information.

Check your understanding

Which contributes more, a shared neighbor with degree 2 or degree 10?

Answer: Degree 2, because its positive logarithm is smaller.

Why can two pairs with two shared neighbors have different scores?

Answer: Their shared neighbors can have different degrees.

What does a high score demonstrate?

Answer: Stronger similarity under this weighting rule, not a confirmed future relationship.

Related terms

Node degree
Link prediction
common neighbours
graph similarity
resource allocation index
Jaccard similarity
recommendation system
information leakage

Learn next

Calculate common-neighbor, Jaccard and Adamic - Adar scores on the same small graph. Then design an evaluation that uses an earlier graph snapshot and clearly defines which later links count as outcomes.

Sources

NetworkX documentation, adamic_adar_index: formula, graph assumptions and implementation scope. https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.link_prediction.adamic_adar_index.html Adamic and Adar, Friends and Neighbors on the Web, author-hosted draft: original research context for inferring relationships from shared information. https://www.cond.org/fnn.pdf The numerical graph, business applications and Cedar Exchange case are original illustrations, not reported research results.

Previous: Adam Next: Adaptive boosting