Skip to content
Back to search results

Bipartite Network Analysis — Two-Mode Networks

Introduced by
Borgatti & Everett (1997) formalised the two-mode network framework
Year
1997
Type
Graph-structural / relational analysis
Font size

What does it do?

Bipartite network analysis, formalised by Borgatti and Everett in 1997, is a graph-structural method for studying networks in which nodes are divided into two disjoint sets — actors and events — and edges exist only between sets, never within them.

It is the natural framework for author–paper, patient–disease, user–product, and any other affiliation data, and it extends one-mode network analysis by providing metrics and projection techniques tailored to the two-mode structure.

The intuition behind the method

Imagine a spreadsheet where rows are people and columns are the clubs they belong to. A cell is marked if person A belongs to club B. Standard network tools expect everyone to connect directly to everyone else, but here people only connect to clubs, not to other people. Bipartite network analysis reads that two-sided structure as it is — measuring how central each person or club is within the whole affiliation pattern — and can then fold it down into a people-only network where two people are linked if they share at least one club, giving a richer picture than simply ignoring the two-mode nature of the data.

History

The formal treatment of two-mode network data was established by Borgatti and Everett in their 1997 Social Networks paper, which extended standard graph metrics to the bipartite case and introduced the projection framework. Guillaume and Latapy (2006) subsequently showed that many empirical complex networks — from the Internet to biological systems — exhibit bipartite substructure, reinforcing the generality of the approach. The method has since become standard in bibliometrics, recommendation-system research, and computational social science.

When to use it

Use bipartite network analysis whenever your relational data naturally divides into two disjoint node types — actors and events, users and items, patients and diagnoses — and you want to study the structural properties of that affiliation pattern. The method applies to exploratory, relational, and classification purposes and handles binary, continuous, or categorical edge weights. At least 20 nodes in total are recommended for meaningful structural metrics. The core assumption is strict: edges exist only between the two node sets, never within a single set. If that partition does not exist in your data, standard one-mode network analysis is more appropriate.

Limitations and considerations

  • Projection to one-mode networks loses information: two actors may share the same projected edge weight despite very different co-participation patterns.
  • Standard centrality metrics designed for one-mode graphs must be adapted or reinterpreted; naive application produces misleading rankings.
  • Very sparse bipartite graphs (few connections per node) yield unreliable centrality estimates.
  • The method describes structure but does not model causal mechanisms; inferring why actors affiliate requires complementary methods.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Respects the inherent two-sided structure of affiliation data rather than forcing it into a one-mode representation that discards information.
  • Projection into one-mode networks enables downstream analysis with the full toolkit of standard network methods.
  • Applicable across diverse domains — bibliometrics, recommendation systems, epidemiology, ecology, and retail — wherever actor–event data arise.
  • Statistical comparison against the BiRG null model allows significance testing of observed structural patterns.

Applications

  • Bibliometrics: mapping author–paper affiliation networks to identify influential researchers and collaborative communities.
  • Recommendation systems: analysing user–item interaction graphs to surface co-purchase or co-viewing patterns.
  • Epidemiology: studying patient–disease or host–pathogen networks to detect comorbidity clusters.
  • Ecology: modelling plant–pollinator or prey–predator bipartite food webs.
  • Organisational research: examining employee–project or board-member–firm networks to study resource sharing and brokerage.

How it works

How it works

1

The raw affiliation data are assembled into a bipartite graph B with two disjoint node partitions — conventionally called the top set (actors) and the bottom set (events) — and edges drawn only across the partition boundary. The input is typically an incidence matrix or a two-column edge list; no edge is permitted within a single partition, which is the defining structural constraint of a bipartite graph.

2

Before computing metrics the two-mode property is verified: each node must belong to exactly one partition and no within-partition edge may exist. A random two-mode null model — the Bipartite Random Graph (BiRG), which preserves degree sequences — can be generated at this stage to serve as a statistical baseline for later comparisons.

3

Centrality and clustering metrics that are adapted for the bipartite structure are calculated on the full two-mode graph. Standard one-mode betweenness or closeness can be computed but must be interpreted with awareness that paths always alternate between partitions. Bipartite-specific clustering coefficients (such as the redundancy coefficient) are preferable because they respect the structural constraint that actors never connect directly to each other.

4

The two-mode graph is optionally collapsed into one or both one-mode projections. In the actor projection two actors are connected if they share at least one event; edge weights typically encode the count of shared events or a weighted co-participation score. Projection discards information, so weighted projections and comparison against the BiRG null model are recommended to assess which co-participations are statistically unexpected.

5

Centrality rankings, clustering coefficients, and projection structure are interpreted in domain terms — for example, identifying the most influential authors in a co-authorship network, the most attended events in an actor–event network, or the strongest co-purchase pairs in a retail network. Findings should be compared against the null model to distinguish meaningful structural patterns from random fluctuations.

Common pitfalls

  • Treating the bipartite graph as a one-mode graph by ignoring partition membership — this violates the structural assumption and produces nonsensical metrics.
  • Using unweighted projection when events differ greatly in attendance size — a co-participation in a two-person workshop should not carry the same weight as co-attendance at a thousand-person conference; use weighted projections instead.
  • Applying one-mode clustering coefficients without adjustment — because actors never share direct edges, naive clustering coefficients collapse to zero; use a bipartite-appropriate clustering measure.
  • Skipping null-model comparison — without the BiRG baseline it is impossible to judge whether an observed co-participation count is structurally meaningful or simply a product of high degree.

Frequently asked questions

How is a bipartite network different from a regular (one-mode) network?

In a regular network all nodes can potentially connect to any other node. In a bipartite network nodes are divided into two disjoint sets — actors and events — and edges are only permitted across the partition boundary. This constraint reflects the natural structure of affiliation data, where people join groups but groups do not directly join other groups.

What does projection do and when should I use it?

Projection folds the two-mode graph into a one-mode graph by connecting actors who share at least one event (or events that share at least one actor). It is useful when you want to apply standard one-mode metrics, but it discards information. Always use weighted projections and compare against a null model to distinguish meaningful co-participation from random overlap driven by high degree.

Can I use standard centrality measures on a bipartite graph?

Standard measures such as degree, betweenness, and closeness can be computed, but they must be interpreted carefully because all paths alternate between partitions. Bipartite-adapted versions of clustering coefficients are strongly preferred over their one-mode counterparts, which collapse to zero by construction in a bipartite graph.

What is the minimum data size needed?

A minimum of 20 nodes total across both partitions is recommended for structural metrics to be meaningful. Very small bipartite graphs produce highly unstable centrality estimates. If your network is smaller, descriptive summaries of degree and co-participation counts are more informative than global structural metrics.

Related methods

Sources

  1. Borgatti, S.P. & Everett, M.G. (1997). Network Analysis of 2-Mode Data. Social Networks, 19(3), 243-269.

    Open publication reference
    Copied.
  2. Guillaume, J.L. & Latapy, M. (2006). Bipartite Structure of All Complex Networks. Information Processing Letters, 90(5), 215-221.

    Open publication reference
    Copied.