Two-mode Network Analysis
- Introduced by
- Breiger, R. L.
- Year
- 1974
- Type
- Bipartite graph analysis
- Family
- Network science
What does it do?
Two-mode network analysis examines networks built from two distinct types of nodes — such as actors and events, authors and papers, or companies and board members — connected only across types.
By analysing this bipartite structure directly or projecting it onto one-mode networks, researchers uncover affiliation patterns, shared memberships, and structural duality that are invisible in standard one-mode social network analysis.
The intuition behind the method
Most social networks involve two kinds of entities: people and the groups, events, or artefacts they belong to. Treating these as a single node set forces false connections between, say, two authors who never co-authored but both published in the same journal. Two-mode analysis keeps both sets separate, respecting their true duality. The sociologist Ronald Breiger formalised this insight in 1974: membership ties simultaneously define person-to-person co-affiliation and group-to-group co-membership, two complementary perspectives on the same data.
History
Ronald Breiger's 1974 paper 'The Duality of Persons and Groups' established the theoretical foundation for two-mode network analysis by showing that actor co-membership and group co-membership are two complementary views of a single affiliation structure. Borgatti and Everett formalised the analytical toolkit in their 1997 Social Networks paper, extending centrality and cohesion measures to bipartite graphs. The approach gained wide empirical use in studies of corporate board interlocks, scientific collaboration, and online community participation, and is now routinely implemented in software packages such as UCINET, igraph, and NetworkX.
When to use it
Use two-mode network analysis when your data naturally record affiliations between two distinct types of entities: authors and publications, patients and diagnoses, directors and corporate boards, or words and documents. It is the appropriate framework whenever you want to understand how membership in common groups shapes actor similarity, or how actors shape group overlap. Do not force two-mode data into a one-mode framework without acknowledging projection artefacts. Avoid it when the two node sets are not conceptually distinct, when the incidence matrix is extremely sparse (most actors affiliated with only one event), or when the research question concerns relations between actors that are not mediated by shared membership.
Limitations and considerations
- One-mode projections introduce edge inflation: large, popular events create cliques among all their members, distorting density and centrality.
- Many standard SNA software routines assume one-mode input and require adaptation or workarounds for bipartite data.
- Interpreting simultaneous actor and event communities requires domain knowledge to label and validate detected groups.
- Sparse incidence matrices (actors affiliated with very few events) produce fragmented projections with limited analytical leverage.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Preserves the genuine duality of actors and groups without collapsing the structure into a lossy one-mode projection.
- Reveals structural positions — such as brokers between otherwise separate groups — that are invisible in one-mode analysis.
- Directly applicable to ubiquitous affiliation data: authorship, board interlocks, event co-attendance, tag co-occurrence.
- Bipartite-native metrics (e.g., bipartite clustering coefficient) avoid the inflated density artefacts of projections.
- Theoretically grounded in Breiger's duality principle, which provides clear interpretive guidance.
- Compatible with weighted and directed extensions for richer affiliation data.
Applications
- Scientific co-authorship networks derived from author-by-paper incidence matrices.
- Corporate governance research: director-by-board interlock networks.
- Political affiliation studies: legislator-by-committee or politician-by-donor networks.
- Online communities: user-by-forum or user-by-hashtag participation networks.
- Bibliometric analysis: keyword-by-document or journal-by-article affiliation matrices.
- Public health: patient-by-diagnosis or contact-by-location exposure networks.
How it works
How it works
The raw data are encoded as a binary or weighted incidence matrix B of size (actors × events/groups), where an entry equals one (or its weight) when an actor belongs to the corresponding event or group. No edges are drawn between two actors or two events; all ties cross the type boundary, producing a bipartite graph.
To apply standard SNA measures, the bipartite graph is projected onto one or both node sets. Multiplying B by its transpose yields an actor-by-actor co-affiliation matrix whose entry A_{ij} counts how many groups actors i and j share. An analogous projection yields a group-by-group overlap matrix. These one-mode projections enable familiar centrality and clustering computations.
Formula
A_{ij} = \sum_{k} B_{ik} B_{jk}Degree, betweenness, and closeness centrality are computed either directly on the bipartite graph (using algorithms adapted for bipartite topology) or on the projected one-mode graphs. Bipartite-native centrality measures avoid the distortions introduced by projection, such as artificially inflated densities in large events.
Community detection algorithms adapted for bipartite graphs — such as bipartite modularity maximisation — identify cohesive subsets of actors and events simultaneously. Alternatively, classic algorithms such as Louvain can be applied to the projected one-mode network, though with some structural information lost.
The observed bipartite network is compared against null models — typically a configuration model that preserves the degree sequence of both node sets — to determine whether detected communities, centrality rankings, or clustering patterns exceed chance. Statistical significance can be assessed via permutation tests or analytical approximations.
Common pitfalls
- Projecting to one mode and then applying standard SNA without disclosing or correcting for projection artefacts.
- Treating the co-affiliation count as a direct measure of tie strength without normalising for event size (Jaccard or cosine normalisation is preferred).
- Using unweighted projections when affiliation frequency or intensity data are available, discarding meaningful variation.
- Applying modularity algorithms designed for one-mode networks to a bipartite graph without using bipartite-appropriate versions.
- Ignoring the null model: reporting observed clustering or centrality without comparing against a degree-sequence-preserving random baseline.
Frequently asked questions
Should I analyse the bipartite graph directly or project it first?
Direct bipartite analysis preserves more structural information and avoids projection artefacts. Projection to one mode is convenient for applying familiar SNA tools but inflates density and can create spurious cliques among members of large events. When feasible, use bipartite-native algorithms and report projection results only as a supplementary check.
How do I handle event size heterogeneity in projections?
Large events generate edges among all their members, artificially boosting co-affiliation counts. Normalise the projected matrix using the Jaccard index (shared affiliations divided by the union) or cosine similarity before computing centrality, to correct for differences in event popularity.
What software supports two-mode network analysis?
igraph (R and Python), NetworkX (Python), UCINET, and Gephi all support bipartite graphs. The igraph bipartite.projection() function and NetworkX's bipartite module handle projection and bipartite-specific metrics natively.
Can I apply community detection to a two-mode network?
Yes, but use algorithms designed for bipartite graphs, such as Barber's bipartite modularity, rather than standard Louvain or Girvan-Newman, which assume one-mode structure. Bipartite community detection assigns co-community memberships to both actor and event sets simultaneously.
Is two-mode analysis appropriate for very small datasets?
With fewer than around 20 actors or events, centrality rankings and community partitions become highly sensitive to individual data points. Permutation-based significance tests are especially important at small sizes, and results should be interpreted cautiously.
Related methods
Sources
Breiger, R. L. (1974). The duality of persons and groups. Social Forces, 53(2), 181–190.
Open DOI referenceCopied.Borgatti, S. P., & Everett, M. G. (1997). Network analysis of 2-mode data. Social Networks, 19(3), 243–269.
Open DOI referenceCopied.